Your own AI
Bring the AI you already pay for
Cygtree runs an AI coding agent you already have — Claude Code, Codex, Cursor, or any Agent Client Protocol agent — as an ordinary program on your own machine, and hands it tools to edit your board. There is no API key field, no cloud relay and no metering. The agent runs as itself, signed into your account, on the plan you already pay for; Cygtree takes no cut of it and adds no limit of its own.
Windows 10 and 11, 64-bit. That is the only build there is. macOS and Linux are hopes, not promises. Cygtree is $20 once, and the AI features are part of the app — there is nothing further to subscribe to here.
If you have no AI tool installed, none of this applies to you and the app still works. Cygtree is a mind map first: cards, lines, images, undo, export. The agent is a thing you can plug in, not the thing you bought.
Which agents work, and how each one is started
Claude Code, Codex, Cursor, any Agent Client Protocol agent, and any MCP client you already use. Cygtree does not wrap an agent or re-implement one: it starts the program you already installed, in the way that program expects to be started, and speaks to it over your own machine's loopback.
| Agent | How Cygtree launches it | What you need installed | What you pay |
|---|---|---|---|
| Claude Code | As a long-running local program that stays open between messages, with Cygtree registered as its only tool server. | The claude CLI, already signed in. Cygtree looks for the native install under your user folder, the npm shim under %APPDATA%\npm, then your PATH. |
Your existing Claude plan. Nothing to Cygtree. |
| Codex | As a long-running local program, started in its app-server mode, sandboxed read-only with its shell tool switched off. | The codex CLI, already signed in. Cygtree looks in the Codex desktop install location, then your PATH. |
Your existing Codex plan. Nothing to Cygtree. |
| Cursor | Fresh for every single turn — no process exists between messages. Continuity comes from resuming the chat, not from a live child. | The cursor-agent CLI, already signed in. |
Your existing Cursor plan. Nothing to Cygtree. |
| Any Agent Client Protocol agent Gemini CLI, Qwen Code, OpenCode, Kilo Code, Cline, Factory Droid, Junie, GitHub Copilot, Cursor CLI, plus Claude and Codex through their own ACP adapters |
You give it a program name and its arguments; Cygtree runs that and talks Agent Client Protocol version 1. There are eleven presets that fill in those two boxes for you. | Whichever agent you picked. Three of the presets run through npx and will fetch their package the first time. |
Whatever that agent already charges you. |
| An MCP client you already use Claude Desktop, another Claude Code, anything that speaks MCP |
Nothing is launched. The client registers one local address once, and you paste it an invite from the board. | The client itself. It does not have to be on the same board, or even be a coding tool. | Whatever that client already charges you. |
You choose the model per agent, and the effort level where the model has one — Cursor takes a mode and a permission level instead. Cygtree refuses to connect on a knob you never answered rather than picking for you; a model with no effort axis is the one case where blank is itself the answer. The model list is read off your own machine — it is not a list shipped inside the app — so a model you got access to this morning is selectable this morning.
What the agent can actually do to your board
Cygtree publishes 38 things an agent can do to a board — read it, build on it, colour it, put pictures on it, remember things about it. That is the part that makes it a board tool rather than a chat window parked next to one. An in-app agent is offered 39 tools, an invited or ACP agent 44 — same board verbs either way, plus the handful of extras an outside client needs to join and to talk back.
Reading what is there
Pull the whole board as a compact tree, or one card in full, or one branch and everything under it. Search cards by text, ignoring case and accents, and get the path back up to the root with each hit.
Building and rearranging
Add a child card or start a new tree. Rewrite a card's text, colour, notes, status, due date, tags. Re-hang a branch under a different parent — it refuses to make a loop. Move, size and pin cards by coordinate. Fold a section away. Delete a branch, undoably. Draw boxes, arrows and lines.
Colour and shape
Colour the cables from one card to its direct children, or give every branch under a card its own hue and let it run downward. Tidy the board, or one branch, into a layout — and the board remembers which one.
Pictures, PDFs, audio, screenshots
Put an image, clip, PDF or a YouTube link on the board from a web address, or a file already on your disk, which gets copied in rather than linked. It can also screenshot the canvas — the whole board, or just what you selected — and drop that on as a card.
Memory that outlives the chat
Notes it can attach to one card, one branch, or the whole board, and read back next time. They live in the board, so the next session — or the next agent, or a different agent — picks up where this one stopped. It can delete one by name too.
Showing you things on your screen
Set your actual selection. Light up cards with a named highlight, or run a timed sequence of them — look here, now here — then clear them. Zoom the camera to a card, or out until everything fits. Open a side panel. Post a one-way notice into your chat window.
Voice, if you have it
Read out which speakers and dictation packs are installed on your machine, and change the speaker, the speed or the read-aloud cap. It refuses a value it cannot honour rather than quietly rounding it to something else.
One undo per job, not per keystroke
A batch of edits can be committed as a single change, so twelve cards appearing is one press of Ctrl+Z, not twelve. Every verb can also carry a revision number, so two agents editing at once cannot silently overwrite each other.
An invited client gets five more. Two are for joining — claiming an invite, and coming back later with a returning key. The other three are the ones deliberately withheld from an in-app agent: it can park and wait for you to type in the board's chat, reply into that chat, and list or switch boards. An assistant draining that queue would be answering messages you meant for somebody else.
What it does not do, including the awkward parts
Four things are simply absent, and there is no setting that turns them on. Then there are four limits that are real.
- No API key field. There is nowhere to put one, because nothing in the app would send it. Cygtree holds no key, resells no tokens and operates no AI service.
- No cloud relay. Every AI path is a local program and a loopback socket on 127.0.0.1. Your prompts and your board do not reach a Cygnus Crane server, because on this path there is no Cygnus Crane server.
- No telemetry in the software. No analytics library of any kind is in the app. Errors go to a log file on your own disk. (This website does measure visits, with a cookieless Cloudflare beacon — that is the site, not the app, and the privacy policy says so.)
- No metering. There is no daily allowance of tool calls, because there is no server in the middle that could count them.
One honest exception to "nothing leaves the machine": licensing. Activating a paid key contacts cygtree.com and our reseller once. After that, that licence verifies offline, permanently, and the site is never contacted again. A trial key is the exception and carries its own separate three-day offline allowance, so it does check back. That is the licence, not the AI — but a page claiming nothing ever leaves your computer would be wrong.
The local connection is not a security boundary
It is a socket bound to loopback — the in-app lane on an ephemeral port,
and the door an outside client registers on a fixed one,
127.0.0.1:8791 — with a token that never touches your disk
and never appears in a command line. That is enough to keep it off your
network. It is not a wall between programs already running as you on your
own computer.
A connected agent brings its own tools, and Cygtree does not sandbox them
Cygtree controls what an agent can do through Cygtree. It does not control what that agent can do on its own. How much it can narrow that differs by agent, and the difference is real — see the row below that matches yours.
The prompts are pre-answered on purpose
An agent running with no terminal cannot answer an approval prompt, so a prompt is a hang rather than a safeguard. Every lane is started with approvals already granted. What bounds the agent is the tool list it was handed, not a dialog nobody can see.
Disk and download reach are off until you turn them on
Out of the box, an agent cannot put a file from your disk on the board and cannot download one. Both switches are yours and both start off. A third — hanging a web link on a card as an attachment — ships on, and lives in the same place if you want it closed. All three live on your machine, never in the board file — so a board somebody emails you cannot hand its author's assistant your disk.
The isolation is not the same for all four, and here is where it differs
| Agent | What Cygtree switches off | What it cannot switch off |
|---|---|---|
| Claude Code | 38 of its own built-in tools — 36 if you switch web access on, which hands it back WebSearch and WebFetch — including Bash, PowerShell, Read, Write, Edit, Glob, Grep and its sub-agents. Cygtree is the only tool server in reach — your other MCP servers are not loaded. | Nothing material. This is the tightest of the four. |
| Codex | Its shell tool, image viewing, hooks, remote plugins and multi-agent mode. The session is read-only and leaves nothing behind. | Codex has no equivalent of a strict server list, so your own configured MCP servers load alongside Cygtree's. If you have some, the agent has them here too. |
| Cursor | In ask mode, file writes and shell are blocked — measured, not assumed. It runs in a private scratch folder, never in your project. | You choose the mode. Pick a freer one and you have chosen a freer agent. |
| Any ACP agent | Cygtree declines to read files, write files or open a terminal on the agent's behalf. | The agent's own file and shell tools, which run inside the agent's process on your disk. This containment is strictly weaker than Claude Code's or Codex's. Bring an agent you trust. |
Cygtree also cannot sign you in. Signing into a coding CLI wants a real terminal, so if the agent says you are signed out, you sign in the way that agent expects and come back. See the troubleshooting section.
What this costs, stated exactly
Nothing beyond the $20. Cygtree adds no AI charge and no cap of its own — the comparison below is why that is worth saying. Miro added an AI connector and meters it: Miro's own documentation publishes the numbers for their MCP server, a daily allowance of tool calls, by plan, after which calls are blocked and you are shown an upgrade.
| Product and plan | Daily cap on AI tool calls | Who imposes it |
|---|---|---|
| Miro, Free | 100 | Miro's MCP server |
| Miro, Starter | 500 | Miro's MCP server |
| Miro, Business | 2,000 | Miro's MCP server |
| Miro, Enterprise | 10,000 | Miro's MCP server |
| Cygtree, $20 once | None of its own | Nobody — there is no Cygtree server in the path to do the counting |
Miro counts every call an agent makes to one of their MCP tools against that daily allowance, whether or not the tool uses their AI credits. Those figures are theirs, published by them, and they can change them.
The boundary, before you get the wrong idea. This is not unlimited AI and nobody here is going to call it that. Your own AI provider's limits still apply in full — the weekly cap on your Claude plan, whatever Codex or Cursor allow you, the rate limits, all of it. What Cygtree does is add no limit of its own on top of the one you already pay for. That is a smaller claim than "unlimited".
What ACP is, and which ACP this is
ACP here means the Agent Client Protocol — the open protocol that came out of Zed, which lets any editor or app talk to any coding agent over a standard wire instead of a bespoke integration per agent. Cygtree speaks version 1 of it.
There is a second thing called ACP and it is not this one. IBM's Agent Communication Protocol is a different protocol, for a different job — agents talking to each other — and it has been archived. If you are comparing the two: Agent Client Protocol is the live one, it is what Zed, and now Cygtree, implement, and it is the one that matters to somebody trying to plug a coding agent into an app.
Why an open protocol is the whole point here
Cygtree's ACP support contains no agent-specific code at all. You type a program name and its arguments, and that is the entire configuration. The eleven presets — Gemini CLI, Qwen Code, OpenCode, Kilo Code, Cline, Factory Droid, Junie, GitHub Copilot, Cursor CLI, and Claude and Codex through their adapters — are nothing more than those two boxes filled in for you. An agent nobody has written yet can drive a Cygtree board on the day it ships, without us releasing anything.
How ACP relates to MCP
They are not rivals. MCP is how a board hands tools to an agent; ACP is how an app and an agent hold a conversation. Cygtree uses both: it speaks ACP to the agent, and inside that session it registers itself as an MCP server so the agent can reach the 38 board verbs. An agent that only speaks MCP — Claude Desktop, say — can still be invited, and gets the same board verbs by a different door.
One thing ACP version 1 does not have
There is no way to steer a turn once it has started. You can cancel it, on every agent. Mid-turn steering — typing a correction while it is still working — exists only on the Codex seat, because Codex's own protocol has it. Nobody else does, and no draft of ACP adds it yet.
When it does not connect
Five failures cover nearly all of it. Four are about the agent — where it is installed, or whether it is signed in — and the fifth is ours.
"We could not find that program"
Set the environment variable for that agent to the full path of the
executable, then restart Cygtree. Cygtree searches a short, specific list of
places and your PATH — it does not go hunting across your
disk — so a CLI installed somewhere unusual has to be pointed at. On
Windows the CLI is often a .cmd shim rather than an
.exe, which is handled, but only if the path you gave points at
the shim that actually exists.
It connects, then says you are signed out
Sign in from a real terminal — claude for Claude Code,
codex login for Codex, cursor-agent login for
Cursor — then come back and connect. Cygtree cannot sign you in, because
the sign-in flow wants a real terminal and there is not one here.
Cursor worked yesterday and cannot be found today
Point at the command name, not at a versioned folder. Cursor installs itself into a folder named after its version and deletes the whole folder when it updates, so a path that pointed at a specific version rots the next time Cursor updates. Cygtree deliberately never stores that versioned path for the same reason.
The model list is empty
An empty list means the agent has not been run and signed in yet on this computer — run it once in a terminal, then reconnect. The list is read off your machine, from the agent's own files: nothing is shipped in the app, and there is no built-in fallback list on purpose, so a model list you see is one your machine really has.
You pressed Disconnect and the agent is still running
End it in Task Manager. An ACP agent you launched directly stops when you
disconnect, but an agent launched through npx — which includes
several of the presets — leaves the real process running behind the
wrapper, because on Windows there is nothing tying the two together. This is
a known limitation.
Still stuck? Write to [email protected]. We will try to reach back out within two business weeks — "try" is the honest word for what one small company can promise.
Questions
Which AI agents can edit a Cygtree board?
Claude Code, Codex and Cursor run as local programs Cygtree starts. Any Agent Client Protocol agent can be used by giving Cygtree the program name and its arguments, with presets for Gemini CLI, Qwen Code, OpenCode, Kilo Code, Cline, Factory Droid, Junie, GitHub Copilot and Cursor CLI. Any MCP client, such as Claude Desktop, can also be invited to a board.
Do I need an API key?
No. There is no field for one anywhere in Cygtree, and nothing in the app would send it if there were. You are already signed into your agent; Cygtree starts that agent and it uses the account it already has.
Does my board or my prompt go to a Cygtree server?
No — on the AI path there is no Cygtree server at all. The agent is a program on your machine and it reaches the board over a socket on 127.0.0.1. Your boards are ordinary files on your own disk, with no cloud, no account and no sync. The one thing that does contact us is licence activation, once.
Is there a limit on how much AI I can use?
Cygtree adds no limit of its own — it has no meter, because it has no server to meter with. Your own AI provider's limits apply exactly as they always did.
What is the difference between Agent Client Protocol and Agent Communication Protocol?
They are two unrelated protocols that share an initialism. Agent Client Protocol, from Zed, connects an app to a coding agent and is what Cygtree implements. IBM's Agent Communication Protocol is about agents talking to one another and has been archived. If a page says "ACP" in the context of plugging an agent into an editor, it means the first one.
Can Claude Code edit my files through Cygtree?
Not through Cygtree — its file and shell tools are switched off in this connection, and Cygtree is the only tool server it can see. That says nothing about Claude Code used normally in your terminal, and for agents connected over ACP the answer is different: those keep their own file and shell tools, which Cygtree does not sandbox.
Does this work on macOS or Linux?
No. Cygtree is Windows 10 and 11, 64-bit, and that is the only build that exists. The other two are hopes, not promises, and nothing on this site should be read as a date.
Do I have to use AI at all?
No. With no agent installed, Cygtree is a mind map and whiteboard — cards, lines, images, PDFs, audio, undo history, search, dictation, read-aloud, export. The AI is a door you can leave shut.
Is the AI included in the $20, or is it extra?
Included. $20 buys the app once, perpetually, and every release of version 1 that follows. Cygtree charges nothing for AI and takes no cut of what you spend with your provider. You can try the whole thing first: the trial runs three days with no payment card. A paid key activates on two computers; a trial key activates on one.