Chats and agents
Start chats with any AI, attach files, pick models, set what an agent may do, and keep track of what's running.
Every chat on the canvas is a full AI agent. It can read and edit files, run commands and build things, in a folder you choose. Run as many as you like side by side, each with its own AI, model and settings.
Start a chat
- Double-click an empty spot on the canvas for a Claude chat (change that in Settings → General), or right-click it and choose New Claude chat from the canvas menu. To use another AI, pick it under Provider in the new chat's model menu.
- Press Ctrl + K and choose New chat in ….
- Double-click inside a project's group, or point at the group's name and click + in its toolbar.
- Or ask Morphy: "open three Codex chats in my website folder" (see Morphy, your assistant).
A chat started inside a project's group works in that project's folder. Anywhere else, or when the project has no folder, it works in Morphy's own folder.
The start screen
A new chat opens on a start screen, with the little Morphy standing in the middle. It says where the chat works ("What are we working on in Website?"), with the model under it, and the branch and how many files are changed when the folder is a Git repository.
- Starters, made from the folder: with uncommitted changes, "Review my uncommitted changes" and "Write a commit message"; on a branch, "What changed on this branch?"; otherwise, "Give me a tour of this project". Click one to put it in the box. Nothing is sent until you send it.
- Pick up where you left off lists the newest conversations in this folder. Click one to continue it here.
- After New conversation, Back to "…" takes you straight back to the conversation you just left.
- Morphy's own chat greets you by name and suggests cards to make.
A small chat shows less: the list goes first, then the line under the question. The start screen goes away with your first message, and Morphy hops down onto the box.
The title bar
From left to right:
- The title. Made from what you asked at first. In a Claude chat it becomes the conversation's name once Claude has given it one, after its first reply. Hover it to see it in full. To name a chat yourself, right-click its title bar and choose Rename…, type the name and press Enter (Esc keeps the old one). Your name stays, also when the chat starts a new conversation, and shows on the phone and in search too. In a Claude, Codex, OpenRouter, Ollama or custom-provider chat, the conversation it's on takes the name as well, so Conversations,
claude --resumeand Codex's own list show it too. That doesn't happen while the chat is in terminal view or works on another computer, and other AIs keep their own name for the conversation. Ctrl + Z undoes it (the conversation gets its old name back too), and Use the conversation's name in the same menu goes back to the automatic title. - The folder the chat works in: the folder's name, or Morphy for Morphy's own folder. Click it to open the folder in your code editor, and more. See Working folder.
- On hover: Close.
- Right-click the title bar for Conversations… (earlier conversations in this folder), New conversation, Rename…, Icon, Standing instructions… and Move to.
An icon and standing instructions
Give each agent its own role. Right-click a chat's title bar:
- Icon › one of the sixteen, Your own… (type any emoji where the name is), or None. It sits before the chat's name.
- Standing instructions… opens a box in the chat, under its title bar: how this chat works, in your words: "You review code for this project. Be blunt and short. Never push to git." The agent follows them in every conversation, whatever its AI. They go along with the first message of each conversation, and once more after you change them. The chat never shows them in its messages.
Or ask Morphy: "make my Codex chat a strict reviewer", "give my chats icons". It sets them, and a line at the foot of the screen says what changed, with Undo. Morphy's own chat keeps its name; its instructions are your AGENTS.md.
Switch to another AI
Click the model button under the text box, point at Provider and pick another AI, for example Codex. AIs that aren't set up yet are marked, with a pointer to Settings → Account; your custom providers are listed there too. The chat keeps showing everything so far, and a line under the last message says what happens next:
Switched to Codex · your next message brings this conversation along (about 2.4k tokens) · Start fresh instead
- Nothing is sent until you send. Your next message takes the conversation along so the new AI can carry on from it: the newest 24 messages at most. The line says about how many tokens that adds to that one message, and once it's gone, it says so ("it got this conversation").
- Switch back and the first AI picks up its own conversation where it left off, and only gets what was said since ("your next message brings what Codex did since").
- Start fresh instead starts a new conversation with the new AI, without the earlier one.
- The chat shows the whole conversation, with a line wherever it changed AI, also after you restart Morphy.
- New conversation, picking another conversation in Conversations, or moving the chat to another folder starts over without the earlier part.
- The files stay in the chat's folder, so the new AI looks at them as they are now.
- Copilot, OpenCode, Hermes Agent and Junie keep their conversations to themselves: what was said with them shows, and goes along, only until you quit Morphy.
Send a message
Type in the box at the bottom and press Enter. Shift + Enter starts a new line, and ↑ at the start of the box brings back what you sent before.
With Shipped work and suggestions on (it's off until you turn it on), the empty box often suggests after a reply what you'll most likely send next, greyed out, like "commit and push it". Press Tab (or click the Tab chip beside it) to put it in the box, then send it as it is or change it first. Start typing and it steps aside. The agent that just answered writes the suggestion at the end of its reply, so it costs no extra AI call. Turn it on or off with Shipped work and suggestions in Settings → General.
Or talk: click the microphone beside the send button, or hold Ctrl + Shift + Space, and what you say is typed into the box. It runs on your own computer; see Voice typing.
While the agent works, everything it does on the way folds into one line under your message ("12 steps · editing app.tsx"): its steps, what it says between them, its thinking, the subagents that finished and the permissions you answered. Before its first step, what it says shows as it comes. After that, a short message goes straight into the fold, and the fold's line shows it while it's being written (or Thinking). A longer message shows under the fold as it comes, since that's most likely the answer. The words it ends on stay under the fold: that's its answer. If its last step came after its last words, those words come out under the fold anyway. When it's done, the line says what's inside ("12 steps · 2 messages"). A reply with no steps shows as it is, with no fold. Click the line to see all of it in order: every file it read or changed, every command it ran and what it said along the way. Click a step to see what went in (the command, the file, a tool's arguments) and what came back (a command's output, a tool's answer), with a button to copy it. A step still running has a pulsing dot, and one that failed has a red one. A command's output shows when it ends, not while it runs, so a command still running after twenty seconds says so when you open it; for a long script, run it yourself in the terminal and let the agent read the log. A command has a button to copy it. Long output shows its first 16 KB, and a chat keeps the output of its newest 200 steps. After a restart a chat shows its steps and subagents again, each with what it gave back (the first 3 KB), in Claude, Codex and Grok chats. Click the file name in a step ("reading warm.png") to open the file in the file viewer, even when it sits outside the project's folder. A timer shows how long it's been working. Code blocks in replies have a copy button.
When Claude hands part of the work to a subagent, the subagent gets a line of its own in the chat, like "Explore · Find the auth check Reading auth.ts 3 steps · 1m 42s": which agent, what it's doing right now, how many steps it has taken and for how long. While it works, its line sits right above Working, with a pulsing dot; when Claude's reply ends first (subagents usually work in the background), it stays at the bottom of the chat until the subagent finishes, and Claude reports back by itself. Click the line to open it: the Prompt Claude gave it (with a copy button), every step the subagent takes as it happens (each one opens like any other), and its Report once it's done. When it finishes, its line says Done (or Stopped) and still opens to all of it. If Claude is still working, the line goes into the turn's fold with the steps. Otherwise it stays where it finished. When Claude sends a finished subagent a new message to carry on, even after a restart of Morphy, that new run gets a line of its own at the bottom, the same way: its Prompt is the message Claude sent, and its steps and report go in it. Until every subagent is done the chat counts as working, even though Claude's own reply is over: its dot keeps pulsing, and it stays with the working chats in the sidebar and in Continue, which says what the subagent is doing. A dev server or other command left running in the background doesn't count. While anything runs in the background (a subagent, a dev server), the chat's title bar shows Background. Click it to see what's running and for how long, and click one to jump to it. Subagents show this way in Claude chats.
When it's done, a quiet line under the reply says how long it worked and when it finished ("Worked 2m 05s · 4:11 PM"). At the right end of that line, the copy icon copies the whole reply as Markdown (what it said in the fold too), and the arrow scrolls back up to the message that started it, so a long answer doesn't lose you. When the turn changed files, an undo icon sits between them (Undo a turn). In a Claude or Codex chat, a branch icon sits just before the arrow (Branch from a turn). The copy and arrow icons sit at the end of the "Working" line while the agent is still going; there the copy icon copies the reply so far. Hover the finished line to see how many tokens the agent wrote in that turn ("3.2k tokens"). Hover the number itself for how much the agent read and how much of that came back from the cache. In a Claude chat, the newest finished line also counts down how long Claude keeps the conversation in its cache ("cache 4:12"). A message sent before it runs out is quicker and uses less of your plan, because Claude re-reads the conversation from its cache. Each message starts the clock again: 5 minutes or 1 hour, depending on your account. After that it says "cache expired". The down arrow at the bottom of the chat takes you back to the latest message.
Arabic, Hebrew, Persian and other right-to-left languages read right to left: the text box turns around as soon as you type, and each paragraph, list and table in a reply follows the language it starts in, so Arabic and English can share a reply. Code always stays left to right. The same goes for the phone app.
Reply to part of an answer
Select some words in a reply and a Reply button appears under them. Click it and that part waits above the text box under "Replying to"; type your message and send. The agent gets the part you picked, quoted, followed by your message, so you can ask about one step of a long answer without copying it by hand.
To answer several things at once (say, the questions at the end of a reply), select the next part and click Reply again: each part gets its own "Replying to" card, in the order you picked them, and what you have typed stays as it is. Many cards scroll. Everything goes in one message: all the parts, then your text. Your message shows each part as a quiet line above what you typed. The × on a card drops just that part.
Undo a turn
When a turn changed files in a chat's project folder, its finished line gets a third icon beside the copy icon: ↶ Undo. Click it to see the files the turn changed, then Undo puts them back as they were before it: changed files get their old content, new files go, deleted files come back. It catches every change in the folder during the turn (the agent's edits and what its commands did) on every AI. A file you (or another turn) changed since is left as it is, and the list says so. If another chat was working in the same folder at the same time, the list warns that some changes may be its.
The line then says undone, and the icon becomes ↷ Redo, which puts the turn's changes back. The agent hears about it with your next message, so it doesn't build on work that's gone.
Undo works in any project folder (it doesn't need to be a Git repository or on GitHub), however many files a turn changed. It doesn't work in Morphy's own folder, a whole drive, your home folder, or your Desktop, Documents or Downloads folder itself (also when OneDrive or iCloud keeps them), nor in a cloud drive's own folder: copying those would read, and download, everything in them. A project folder inside any of them works. It needs the Git program, 2.32 or newer, installed on your computer: Morphy keeps the "before" copies in a Git store of its own, never in your repository. With Git older than 2.40 (Apple's own, on many Macs), a project whose .gitattributes changes line endings has no Undo; install a newer Git to get it there. When a turn has no Undo, its icon shows dimmed, and hovering or clicking it says why. Turns stay undoable for a day. Things your project's .gitignore leaves out, node_modules and files over 50 MB aren't covered, and a folder that would take more than 1 GB to copy has no Undo.
Generated output in bulk isn't covered either. When a turn writes a big batch of files of one type into one folder, like a render's frames or a dataset (100 files and 20 MB, or anything past 200 MB in one turn), Undo leaves that batch out from then on. The turn's line says so. The rest of the turn, like the script that made them, stays undoable. The copies also have size limits. A project's copies over 2 GB lose their oldest turns, all copies together stay under 5 GB, and the copies of a folder that was deleted or hasn't been used for a day are removed. Morphy checks this by itself soon after it starts and every hour. The copies live in the .morph/checkpoints folder of Morphy's data folder (%APPDATA%\Morphy on Windows, ~/Library/Application Support/Morphy on a Mac). If you need the space at once, quit Morphy and delete that folder: you only lose Undo on past turns.
Branch from a turn
Want to try another way without losing a chat that's going well? Hover the line under a finished reply ("Worked 2m 05s · 4:11 PM") and click the branch icon at its right end. A new chat opens beside this one, named "Branch ·" and the chat's name, with the caret in its text box. It shows the conversation up to that turn, then a line: Branched from and the chat's name (click the name to go back to that chat). The AI remembers all of it, what its tools gave back too, and carries on from there with your next message. The chat you branched from stays as it was.
It works in Claude and Codex chats, and in the OpenRouter agent, Ollama and custom providers, which run on Codex: the copy is made by Claude Code's and Codex's own tools, so nothing is left out. A chat switched between AIs branches on the AI that answered that turn. An older Codex can't make the copy and says to update it. A chat in a folder on another computer, or open as a terminal, can't be branched yet.
The branch works in the same folder, with the same model and permissions. Only the conversation is copied, not your files: if the chat changed files after the turn you branched from, those changes are still in the folder, and the line under Branched from says so. To try something without touching the folder, give the branch its own copy. In a focus, the branch joins the focus as a tile of its own.
Files named in a reply
When a reply names a file, you can click it: Morphy turns paths into links once the reply is finished: a path in code style (src/app.js, also with a line number like src/app.js:42), a full path written in the text, and a link to a file in the chat's folder. Only files that are really there become links, so nothing that just looks like a path does. A file in the chat's folder opens in the file viewer; one anywhere else opens in its own app, and a folder opens in File Explorer (Finder on a Mac). Programs and scripts ask before they run, as any link does.
Right-click
Right-click in a chat for the usual text menu. On a link: Open, Copy link for a web address, and for a file Show in File Explorer (Finder on a Mac) and Copy path. A file named in a step has the same menu. On selected text in the conversation: Copy, and Reply to this (the same as the Reply button). On a message with nothing selected: Copy message (a reply comes out as Markdown). In the text box: Cut, Copy, Paste and Select all, and in a chat's text box Attach files…. Paste works like Ctrl + V, so a copied picture or copied cards land as attachments. Every text box in Morphy has the same menu.
Attach files and images
- Drag files from File Explorer (Finder on a Mac) onto the chat.
- Paste an image or screenshot straight into the text box (Ctrl + V, or right-click → Paste).
- Right-click the text box → Attach files… to pick files.
The agent gets the files' locations on your computer and reads them itself, so large files and whole folders of context are fine.
Pictures show as small previews above the text box, other files by their names. Point at a preview to see its name; its × takes it off. Once you send, the pictures show in your message.
Click a picture, above the text box or in a message, to see it large. ← and → (or the arrows at the sides) go through the other pictures of that message, and the buttons beside it open it in its own app or show it in File Explorer (Finder on a Mac). Press Esc, click ✕ or click beside the picture to close it. Right-click a preview for the same, and Copy path. A pasted screenshot is kept for a week; after that, its message shows its name instead.
To give a chat what's on some of your cards, select the cards, click Copy these cards as context in the toolbar and paste into the chat. Pasted into another app, they come out as plain text.
Slash commands
Type / in an empty box for a menu of your skills: /skills lists them all, and typing a skill's name finds it. Pick one to attach it to the message.
If you use Claude Code, the menu also lists your own Claude Code skills and your plugins' skills and commands, under Claude Code, in Claude chats and in Morphy's own chat. Typing the start of a name finds it (/gsap-react finds gsap-skills:gsap-react). In a Claude chat, picking one types it the way Claude Code takes it (/gsap-skills:gsap-react), and Claude Code loads the skill. In Morphy's chat, it asks Morphy to use it: Morphy reads the skill and follows it, whichever AI it runs on. See Your Claude Code setup.
It works in the middle of a message too: type / after a space, and the skill you pick is attached while the rest of your words stay as they are. There, Enter still sends the message as you typed it: choose a skill with ↑ ↓ first, or press Tab to take the one highlighted.
The menu doesn't list the AI's built-in commands (Claude Code's /clear, /model and the rest): in a chat many of them don't do what they do in a terminal. To start over, use New conversation (right-click the chat's title bar); to change the model, use the model menu. A command you type yourself still goes to the AI as it is.
Models and effort
The button under the text box shows the chat's model. Click it for the menu, which also holds the model's effort and the chat's permissions. Point at Effort, Permissions, Provider or More models and its choices open beside the menu. With two Claude, Codex or Grok accounts saved, Account picks the one this chat runs on (more). Each chat keeps its own model and effort: picking one in a chat never changes another. A new chat starts on Default and Auto, unless you picked others for new chats of that AI (Use for new … chats, below).
- Models. The model for your next message. The menu shows the newest model of each kind, for example Opus 5.5, Fable 5.1, Sonnet 5.5 and Haiku 5.5, plus the one you picked; the rest are under More models. Claude shows Anthropic's current models; other AIs show the models your account has. The first line, for example Claude's default, is your AI's own default model (Morphy doesn't pick one per message), and once Morphy has seen it run, the menu names it, for example Claude's default · opus 5.5. It isn't the model you picked for new chats: a new chat opens with that one ticked. In OpenRouter chats, the newest free model comes first, then the newest model of each lab.
- Effort. How hard the model thinks before answering, from Low to Maximum for Claude. Claude's last choice, Ultracode, is Claude Code's own: Extra high, and for every real task Claude plans and runs a team of agents (a workflow) instead of working alone. It gives the most thorough answers and uses far more of your plan. It needs Claude Code's workflows on your Claude plan; without them it runs as Extra high. For Codex the menu lists what the chosen model takes (for example Low to Maximum, or up to Ultra), so it changes with the model. Auto lets the model decide; its line names the model's own default. A level the new model doesn't have runs on the closest one it does. For Ollama, see Thinking.
- Permissions. What the agent may do without asking you: see Permissions.
- Mode. Only in a chat on your own ACP agent that has modes of its own, for example one for planning: pick one, or Default for the mode its conversations start in. Such an agent shows Effort only when it offers levels, and models only when it lists them.
- Provider. The AI this chat runs on, your custom providers and ACP agents included. Pick another to switch. Morphy's own chat changes AI in Settings → General instead.
- More models. The other models, and Custom model…: click it, type any model name where it was, and press Enter (Esc leaves it).
- Search models. When an AI has a long list (12 models or more, like OpenRouter's), More models opens with a search box at the top, ready to type in. It searches every model, also the ones shown in the menu itself. Type a few letters of a name, a maker or a heading (for example
gemini,free glmorpaid), press Enter to pick the first one left, or Esc to empty the box. ↓ moves into the list, ↑ and ↓ go between the models, and ↑ from the first goes back to the box. The headings stay at the top while you scroll. - Use for new Claude chats (named after the chat's AI). At the bottom of the menu: from now on, new chats of this AI start on this chat's model and effort. That counts for every new chat, including those from a double-click, New chat buttons, the canvas menu, your phone and the chats Morphy starts. Once they're the ones, the line reads Default for new Claude chats. Chats you already have keep their own. You can see and change them in Settings → General → New chats, where Default and Auto go back to the AI's own. If your sign-in can't use that model (Codex refuses some models on some plans), the chat runs on another one, says so, and new chats of that AI go back to its default model.
- Context. At the top of the menu, once the conversation has an answer, also when you come back to a chat days later: how full the conversation is, for example 62k of 200k tokens, with a bar that turns blue when it's four-fifths full. When it fills up, Claude and Codex compact the conversation on their own; you can also start a New conversation (right-click the chat's title bar). Claude, Codex, OpenRouter and Grok say how big the model's window is, and so does a custom provider whose service gives it; for the others the row shows only what's used, and for Antigravity there's no row.
- The context ring. The same at a glance, without opening the menu: a small ring at the left of the text box's bottom row, next to the model button. It fills up as the conversation does and turns blue at four-fifths; point at it for the numbers. It shows once the AI has said how big the model's window is (Claude, Codex, OpenRouter, Grok, and a custom provider whose service gives it).
Hover the model button to see which model the last turn ran on, how big the conversation's context was and how many tokens the reply took.
Permissions
Permissions in the model menu sets what the agent may do without asking you. Each chat keeps its own setting. You can also change it from your phone, along with the chat's model and effort.
| Mode | Label in Claude chats | Label in other chats | What the agent may do |
|---|---|---|---|
| No guardrails | No guardrails | No guardrails | Edit files and run commands without asking |
| Auto | Auto | – | Claude Code checks each action: routine ones go through, risky ones ask first |
| Edits only | Auto-accept edits | Workspace write | Change files in its folder; anything else asks first |
| Ask first | Ask first | Ask first | Ask before it changes a file or runs a command |
| Read only | Plan only | Read only | Read and plan, but change nothing |
On a new install (from Morphy 0.1.22), new chats start in Auto-accept edits. An install from before keeps the default it had: No guardrails, unless you changed it. Change the default under Permissions in Settings → General → New chats.
When an agent asks, the chat shows what it wants to do with Allow, Always and Deny. Always allows that kind of action for the rest of the session (until you quit Morphy). If you deny, the agent adjusts its approach or asks you.
You can change a chat's permissions while it works. Switch a Codex chat (or Codex on OpenRouter, Ollama or a custom provider) to No guardrails and the reply it's working on goes on without asking: a request already on screen goes through, marked Allowed (No guardrails), and only the commands on your Always ask before list still ask. Other AIs take a new setting with the next message, and so may Codex when the new setting is stricter.
Claude Code keeps a few safety checks of its own, even in No guardrails. For example, it asks before a command that changes folder and then deletes files with a wildcard (cd build; rm -f frames/*), because it can't tell for sure what that would delete. The request shows Claude Code's reason under the command.
Auto
Auto is Claude Code's own mode, for Claude chats. Before each action, Claude Code runs a safety check on it: routine edits and commands go through without asking, and anything that looks risky asks you first, with the reason under it. Sometimes the check refuses an action outright; Claude then tries another way or tells you.
- Not every model or Claude plan has Auto. Haiku doesn't, and it doesn't run while Claude Code's fast mode is on. When Auto isn't available, the chat switches to Ask first and says why. Pick another model, then auto again.
- The safety check runs on Claude too, so Auto can use a little more of your Claude plan than Auto-accept edits.
- Always ask before still asks first in Auto.
- You pick Auto per chat, in the chat's model menu or from your phone. It isn't one of the choices for new chats.
- Other AIs don't have it. A chat in Auto that you switch to another AI moves to Workspace write.
Always ask before
Some commands are hard to undo. Settings → General → Always ask before lists the ones every agent asks you about first, in every mode but Plan only, even No guardrails. It starts empty. Add commands the way you'd type them, like git push, rm -rf, npm publish or git reset --hard.
When an agent is about to run one, the chat shows it with Allow and Deny, and says it's on your list. There's no Always: it asks every time. You can answer from your phone too. Everything else runs as the chat's mode says.
A command counts however it's written: with options added (git push --force), in a chain (cd app && git push), inside PowerShell or bash -c. Extra words after the program have to match in order, so git push doesn't catch git stash push. Words starting with - can be anywhere, so rm -rf also catches rm -r -f.
It works for Claude, Grok, Copilot and OpenCode in every mode but plan only. For Codex, and Codex on OpenRouter, Ollama or a custom provider, it works in No guardrails: in Ask first and Auto-accept edits Codex runs commands inside its own sandbox, limited to the chat's folder, without handing them to Morphy to check. Antigravity asks in No guardrails; in its other modes it can't run commands from Morphy at all. Hermes Agent and Junie decide by themselves which commands to ask about, so your list covers only those; their chats say so once. It doesn't cover a chat opened as a terminal, or a script that runs the command itself.
For Antigravity on Windows, Morphy adds a small plugin to Antigravity's own settings folder, ~/.gemini/config/plugins/morphy-ask-before, the first time an Antigravity chat in No guardrails runs with a command on your list. Antigravity hands it each command before running it, and the plugin checks it with Morphy. Anywhere else, like agy in your own terminal, it does nothing. Delete the folder to remove it; Morphy adds it back the next time it's needed. On a Mac or Linux, Antigravity doesn't check your list yet, and its chat tells you so.
When an agent asks you something
Claude and Codex can stop to ask you a question with a few options to choose from, in any mode. So can Codex's harness on OpenRouter, Ollama and a custom provider. The chat shows each question with its options. Pick one (or several, when Claude says you can), or type your own answer in the field under them, then click Send answer. A single question with one choice to make is sent the moment you click an option. Skip lets the agent choose itself; it says which choice it made, so you can change it.
In Plan only, Claude ends with its plan and asks whether to go ahead. Approve lets it start working. The chat goes back to the permission mode it was in before you switched to plan only, or to the permissions you set for new chats (Settings → General → New chats) if it started in plan only. Keep planning sends it back to the plan. Type what should change in the field above the buttons first, and Claude revises the plan and shows it again.
You can answer questions and plans from your phone too.
No guardrails means (almost) no questions
In No guardrails an agent can change and delete files and run any command in its folder without checking with you. That's what makes it fast. For a folder you can't easily restore, use Ask first, or keep it in Git.
Antigravity can't ask from inside Morphy: in Ask first and Workspace write it skips any step that needs your OK, and tells you so and which permissions that message ran in. In Read only it can read the files in its folder but can't run any command, not even one that only looks, and it says so when it skips one. To let it do everything, set the chat to No guardrails. A chat that a flow or Morphy makes starts with the permissions in Settings → General → New chats, so check its own when its job needs commands.
Stop, send while it works, and schedule
The button next to the text box changes with what the agent is doing:
-
Send when it's idle and there's something to send. With an empty box there's no button, and the microphone shows only while you point at the chat or type in it.
-
Stop this turn while it works and the box is empty. Stop always stops: if an agent doesn't respond, Morphy ends it after a few seconds. The conversation is kept.
-
Send while it works and you've typed something. What happens next depends on the AI:
- Claude and Codex take it the way Claude Code does. Your message goes to the agent right away and waits under the "Working" line, marked next step. The agent reads it at its next step, between two actions, without stopping, and carries on with it in mind; the message then moves up to where it was read.
- Claude: if it finishes before it gets to a next step, your message runs on its own right after.
- Codex: if it's in the middle of writing, it finishes that reply first, then reads your message in the same turn. The OpenRouter agent, Ollama and custom providers run on Codex and work the same way, except that with them Codex stops the reply it was writing and starts again with your message.
- Other AIs (Grok, Copilot, OpenCode, Hermes Agent, Junie, Antigravity) queue it: your message waits, marked queued, and is sent as soon as the current turn finishes.
The waiting message has Send now, which stops the current turn so your message goes right away, and Delete, which takes it back. Claude's can be deleted until Claude reads it; Codex's only until Codex has received it, usually a moment after you send it, when the Delete button disappears. If you stop the turn instead, messages the agent hasn't read yet come back into the box.
- Claude and Codex take it the way Claude Code does. Your message goes to the agent right away and waits under the "Working" line, marked next step. The agent reads it at its next step, between two actions, without stopping, and carries on with it in mind; the message then moves up to where it was read.
Slash commands sent while Claude or Codex works wait for the turn to finish, like a queued message.
Queued messages live as long as Morphy is open; quitting or restarting drops them. A message the agent already read is part of the conversation and stays.
Send later
Type your message, then right-click the send button. Send later offers:
- In 30 minutes, In 1 hour, In 3 hours: each with the time it means.
- When the limit resets: shows up when your Claude, Codex or Antigravity plan is nearly used up. It sends a minute after the limit resets, so a long job carries on by itself.
- Tomorrow morning: 9:00 AM.
- At: pick a time and click Schedule. A time that has already passed today means tomorrow.
Turn on Repeat first to send it again and again: every 30 minutes, hour, 5 hours or day. "continue" every 5 hours keeps an agent going through each new limit window.
The message waits as a row above the text box, with when it goes. Hover it for Send now (↑) or × to remove it. When it goes, it shows in the chat like any message, marked scheduled.
Scheduled messages are kept by Morphy, so they survive closing Morphy or restarting the computer: one that came due while Morphy was closed goes as soon as it opens again. One that comes due while the agent is working, or waiting for your answer, waits ("Due · after this reply") and goes right after. It's sent with the chat's model and settings at that moment.
When you hit a usage limit
When a turn stops because your plan's limit is used up ("You've hit your session limit"), the chat says so under it, with when the limit resets, and these buttons:
- Switch to … shows when you keep another Claude, Codex or Grok account in Morphy that hasn't hit its limit. It switches to that account and sends your message again. Morphy can also do this by itself: see limits and projects.
- Send when it resets schedules your message for a minute after the reset, like Send later. If the turn got nothing done, your message goes again; otherwise "Carry on where you left off". When Morphy doesn't know the reset time, it opens the Send later menu instead.
- Use free reset shows on a Codex chat when OpenAI has given your account a free limit reset (it also shows in Plan usage). Morphy asks first, because a reset can't be undone. Use resets your weekly and 5-hour limits and sends your message again.
- Continue with Codex (or whichever other AI you have connected) opens a new chat beside this one, in the same folder and project, and hands it the conversation so far, so it carries on with your last request. Other AI… lists the rest. The first chat stays as it was.
- Connect another AI shows when you have no other AI connected, and opens Settings → Account.
How each plan's limits work, and how to make them last: Claude Code usage limits and Codex usage limits.
When a turn stops because the account has no credits or plan for that AI ("Credit balance is too low", "Insufficient credits", or a ChatGPT workspace that is "out of credits"), the chat shows the same card: "Claude has no credits or plan for this." That answer comes from the AI's side, not from Morphy, and nothing resets it. If you keep another Claude, Codex or Grok account in Morphy, the card offers Switch to … that account first, as at a usage limit (or Morphy switches by itself, if you turned that on); an API key without credits doesn't switch. The card also offers Continue with another AI you have connected (or Connect another AI).
If an agent goes quiet for 10 minutes, the chat asks whether to Stop or Keep waiting.
On Freemium, 3 agents work at a time (Morphy's own chat doesn't count). A message to a fourth waits, the chat says so, and it starts by itself as soon as one of the others finishes. Stop drops it.
Know when it's done
- The bell. Select a chat and click the bell in the toolbar. When the chat finishes, or stops to ask you something, a chime plays until you click, scroll or type in that chat. To pick another sound, or one of your own, see The bell's sound. While it rings, the chat sits at the top of Continue in the sidebar with a little swinging bell; click it there to go to the chat, which stops the bell.
- Tabs. A focus or canvas tab shows a pulsing dot while one of its agents works, and a brighter one while an agent waits for you. A blue number counts the agents that finished, or need you, while you were looking at another canvas. The Main canvas tab holds every chat, so it only shows the number.
- From far away. Zoomed out, every chat shows Working… or Needs you.
- The sidebar. A project wears the same blue number as the tabs, or a pulsing dot while an agent there works or waits for you. The last line of the sidebar sums it up (1 needs you · 2 working); click a part to go to that chat.
- The little Morphy on the chat's box (below).
Morphy on every chat
The little Morphy stands on the top edge of every chat's box, in the shape of the Morphy logo and in your accent colour. While a chat is quiet, it hides behind its box with just the top of its head and its eyes over the edge. Click into the box, or on it, and it comes back up. It stays up while the agent works, and shows what it's doing:
- Thinking: a thought cloud.
- Reading a file: a book, its eyes reading along. Searching: a magnifying glass. Editing: a pencil, scribbling. Running a command: a little terminal. On the web: a spinning globe.
- Subagents: a mini-Morphy for each one at work, which jumps back in when it's done.
- Needs you: it stands up and waves, then keeps its hand up until you answer.
- Done: it hops and cheers, with confetti after a long turn. Something went wrong: X eyes and a sweat drop.
- Waiting for a rate limit or a free agent: an hourglass. While you dictate: a microphone.
- After ten quiet minutes it falls asleep and melts into a puddle; your next message, or pointing at the chat, wakes it.
Point at it to see what it's doing in words, like "Reading main.js". It watches you type, looks up when you drag files onto the chat, and wobbles when you click it. At rest it stays still, so a canvas full of chats costs nothing, and it pauses while Morphy isn't the window you're using. Zoomed far out it isn't shown. With Animations set to Reduced it still changes shape, without the motion.
In Settings → Personalization → Morphy you can change what it's made of. Crystal is glass, like the logo: a glass book, a crystal ball on the web, and it frosts over and cracks when something goes wrong. Pixel is a little sprite whose pages turn and sand drops pixel by pixel; it glitches when something goes wrong. Ink is drawn like a comic-book hero: it smirks at rest, raises a brow while it searches, puts up a "!" balloon when the chat needs you, gives a thumbs-up when it's done and swears in symbols (#@$!) when something goes wrong. Cloud drifts instead of hopping: it rains while a command runs, the sun breaks through behind it when it's done, and it turns into a little thunderstorm when something goes wrong. Slime jiggles and splats: bubbles drift up in it while it works, it holds up a "!" of goo when the chat needs you, and it melts when something goes wrong. Flame is fire: its flames flicker and embers rise while it works, it flares and throws sparks when it's done, and it sputters red when something goes wrong. Holo is a hologram over a little projector: it glitches into each new shape, a ring of light pings out when it's done, and its picture breaks up when something goes wrong. Clay moves like stop-motion: clay balls pop up while it thinks, pellets of clay bounce out when it's done, and it cracks when something goes wrong. Every style but Classic is part of Pro. You can also change its size, or take it off the box: it then only stands on an empty chat's start screen. See Morphy on the chats.
Working folder
Every chat works in a folder: it reads, edits and runs commands there. You can give it more folders too.
- Open it in your code editor: click the folder in the chat's title bar and choose Open in VS Code (or Cursor, Windsurf, Antigravity IDE, Zed or VSCodium, whichever you have; with several, pick one in Settings). Show in File Explorer (Finder on a Mac, Files on Linux) opens the folder itself.
- Change it: click the folder and choose Change folder…, then pick another one. The chat starts a fresh conversation in the new folder.
- Give it more folders: click the folder and choose Add a folder…, then pick another one, like the backend next to your frontend. The chat keeps its conversation and works in both from its next message, with the same permissions as in its own folder. Its folder then reads like
web +1(hover it to see them all), and the folder's menu lists them under Also works in: click one to take it away. Up to 8, in Claude, Codex, Copilot and Antigravity chats (Grok and OpenCode can't take more folders). The others stay when you change its folder; Switch to Morphy Agent drops them. - Back to Morphy's own folder: click the folder and choose Switch to Morphy Agent. The chat works in Morphy's own folder again, on a fresh conversation, and can make cards and run the canvas like Morphy. Not while it's answering: changing the folder restarts the agent.
- Move it into its project: when the folder (or a folder inside it) belongs to one of your projects and the chat isn't in any project yet, its title bar shows Add to project next to the folder; one click puts it in. The folder's menu starts with the same thing, Add to the … project (or Move into the … project when the chat is in another project). The chat moves into the project's group, and the canvas follows it there. Ctrl + Z puts it back.
A chat in Morphy's own folder shows just a folder icon when you hover it: click it to pick a project folder.
Your Claude Code setup
A Claude chat loads your own Claude Code setup from ~/.claude (%USERPROFILE%\.claude on Windows), as Claude Code does in a terminal: your settings and CLAUDE.md, your skills, and the plugins you turned on, with their skills, commands, agents, MCP servers and hooks. That goes for a chat in a project's folder and for one in Morphy's own folder, like the chats a flow makes. A project's own .claude folder and CLAUDE.md load too.
Morphy's own chat is the exception: it runs your canvas, and your hooks or rules could get in its way, so it loads only its own instructions. It still knows your skills and plugins. Each time Morphy starts, it reads their names and descriptions (never your hooks, your CLAUDE.md or any key) and uses a skill when a task fits it, whichever AI it runs on. When a plugin needs its own tools, MCP servers or hooks, Morphy hands the task to a Claude chat, which has them, and tells you what came back. To add or change skills and plugins, use Claude Code: Morphy only reads them.
Its own copy
Several agents in one folder can overwrite each other's work. To keep one apart, click its folder and choose Work in its own copy (the folder must be in a Git repository with at least one commit). The chat gets its own working copy of the repository (a git worktree) on a new branch, morphy/ + the chat's name, starting from your last commit, and works there in a new conversation. Your folder stays exactly as it is. Its folder reads like my-app · own copy. More on keeping parallel agents apart: Claude Code worktrees and how to run several Claude Code agents at once.
- Review its changes… (in its folder menu) lists every file it changed, with lines added and removed. Open the diff shows the whole change in the file viewer.
- Bring into my-app puts its changes into your folder, not committed and not staged, so you can look them over there. Next time, only what changed since comes in. If you changed the same lines in your folder meanwhile, Morphy says which files clash; Bring in, marking the clashes brings in the rest and marks those lines with conflict markers (
<<<<<<<…>>>>>>>), like a git merge. - Commit to its branch commits everything in its copy to its branch, to merge your own way.
- To stop, click Its own copy · morphy/… in its folder menu. If it has changes you haven't brought in, Morphy asks first. Its copy is deleted, and its branch too unless it has commits; the chat goes back to your folder in a new conversation.
Things Git ignores, like node_modules, aren't in the copy, so the agent may need to install them there; .env files are copied. Copies live in Morphy's data folder, not next to your project. Not while the agent is working, and not for chats on another computer.
In a project folder, a chat is a normal coding agent. A Claude chat there is set up exactly like Claude Code in your terminal: it uses your own Claude Code setup (the .claude folder in your user folder, with your settings, CLAUDE.md, skills, plugins, agents and hooks), plus the project's own CLAUDE.md and .claude folder.
In Morphy's own folder, a chat gets Morphy's instructions instead, so it can build cards and manage other chats. Your personal Claude Code setup isn't loaded there, so your own hooks, plugins or rules can't get in the way of how Morphy runs the canvas.
A folder on another computer
A Claude or Codex chat can also work in a folder on another computer you reach with SSH: a Raspberry Pi, a server, a desktop at home (over Tailscale or the same Wi-Fi). Turn on Folders on other computers in Settings → General → Agents first.
- Click the chat's folder and choose Use a folder on another computer…. A terminal opens on the right.
- Computer: type its user and address, like
you@100.91.156.113(or thessh …line you'd type in a terminal), then Enter. - The first time only, type that computer's password. Morphy leaves a key of its own there, so it never asks again. The password isn't saved.
- Folder: the folder there, like
~/myapp. If it isn't there yet, Morphy offers to make it. - If Claude Code or Codex isn't installed or signed in on that computer, Morphy offers to do it there. The sign-in page opens in your browser here; for Claude, paste the code it shows back into the terminal.
The terminal says Done and closes by itself. From then on the chat works on that computer, and its folder reads like myapp · jarfish-pi; everything else about the chat stays the same. If the computer is off or out of reach when you send a message, the chat says so, and your next message tries again. Ctrl + C in the terminal calls the setup off.
- The other computer needs Linux or macOS, with SSH turned on. On Windows, Morphy uses Windows' own SSH client (if it's missing: Settings → System → Optional features → OpenSSH Client).
- The conversation lives on that computer, like any Claude Code or Codex session there; Morphy keeps a copy of each reply to show it again.
- Only Claude and Codex chats can work there, and they have no terminal view yet.
- To come back, click the folder and choose Choose a folder on this computer… or Switch to Morphy Agent.
- Turning the setting off keeps chats already set up working; it only hides the menu item.
Conversations
Chats remember their conversation. Close Morphy, open it tomorrow, and every chat continues where it left off.
- New conversation (right-click the title bar) starts fresh. Files and cards are untouched. Its start screen offers Back to "…", in case you want the last one back.
- Conversations… (right-click the title bar) lists earlier conversations in this chat's folder, so you can pick one up again. Each one shows its name, the same name Claude Code and Codex show: one you gave it (renaming the chat in Morphy,
/renamein Claude Code, or renaming a thread in Codex), or else the name Claude gave it. A conversation without a name shows its first message. - A Codex conversation can only be continued in one place at a time. One that another chat or Codex itself (its app, VS Code or a terminal) has open waits until it's closed there: the chat says so, doesn't send your message, and stays on its conversation (more).
Open a chat as a terminal
Prefer your AI's own command-line screen? Select a chat and click Open as a terminal in the toolbar (the icon next to the bell). The card then runs the AI's own app: Claude Code, Codex, Grok, Copilot, OpenCode or Antigravity, with its own welcome screen and logo. It stays in the same folder and on the same conversation, with the chat's model, effort and permission mode. (OpenCode 2 opens the conversation on the model and mode it was last on.) OpenRouter, Ollama and custom-provider chats open Codex, which those chats run on. When that app isn't installed on your computer (or a custom provider isn't set up yet), the button is greyed out; point at it to see why. It works while signed out too: the app then asks you to sign in.
- Type in it like any terminal. Esc goes to the app (in Claude Code it interrupts). Ctrl + C copies selected text, and otherwise goes to the app. Ctrl + V pastes text. Right-click copies selected text, or pastes. On a Mac, use ⌘ + C and ⌘ + V.
- Send a screenshot or a file: paste a screenshot or a copied picture with Ctrl + V, or drop files on the card. Morphy saves a pasted picture as a file and pastes its path, as a terminal does with a dropped file. Claude Code and Codex attach an image path as an image ([Image #1]); other apps get the path to read.
- Scroll in it when the card is selected; otherwise scrolling moves the canvas, as with any chat.
- Zoom changes its text size: the canvas zoom, or the chat's own zoom (− and + in the toolbar).
- The dot in the title bar pulses while the app is working. Once it's done, the chat moves to the top of Continue, as after any reply, on your phone too.
- Back to the chat: click the button again. The chat continues the conversation you were in last in the terminal, including one you started there (for example with
/clear). A conversation you only reopened there (/resume) isn't followed: the chat goes back to its own. Conversations from elsewhere in the same folder, like Claude Code in VS Code or another chat's terminal, are never taken. - If the app ends (
/exit), the card says so: press Enter or click Start again, or go Back to chat.
New chats as a terminal: to have every new chat of an AI start like this, set Opens as to Terminal in Settings → General → New chats (it's per AI: Claude can open as Claude Code while Codex chats stay chats). Shift + double-click on the canvas makes the other kind: a terminal when new chats open as chats, a chat when they open as terminals. A chat that starts with a message, like a flow's or one from your phone, is always a chat.
A chat stays a terminal until you switch it back, even after Morphy restarts. Switching only works while the agent is idle. While a chat is a terminal, it takes no messages from your phone or from Morphy. Changing its folder starts the app again in the new folder, on a new conversation. In a folder it hasn't seen before, the app may first ask whether you trust it, as it would in any terminal. Morphy's own chat stays a chat. On Freemium, a chat open as a terminal counts as one of your three agents at a time.
Claude Code comes with Morphy. The other apps are the same ones their chats need (see Connect your AI). For a plain terminal, use the dock's Terminal.
Close and bring back
Close a chat with its ✕. A chat can't be closed while its agent is working: stop it first. Ctrl + Z brings a closed chat back on the same conversation, and for 3 days you can restore it from Ctrl + K.
If agents are still working when you quit Morphy, it lists them and asks: Keep working or Quit anyway. Your chats and their conversations are kept either way.
Chats working together
To have chats hand their work on to each other, share a list of tasks, or answer one question from several sides, put them in a flow (Pro).
Claude chats can also message each other directly, with Claude Code's own messaging: ask one to send something to another. Each Claude chat keeps one name for this, its folder's name and four letters or digits (Morphy's own chat is morphy). A chat that sends or gets such a message stays ready for two hours after the last one. A chat you haven't used for a few minutes doesn't show up for the others until you send it a message.
