Unlimitless
保存你已定下的決策及其原因，並在你接下來所使用的任何 AI 介面中自動呈現。適用於 Claude、Claude Code、ChatGPT、Cursor、Cline、Kilo Code、opencode、Copilot、Antigravity 與 Devin Desktop。
文件
Unl
Unl joins the AI you already use
Your why agent keeps what you decided and why, and hands your agent the part that bears before it acts. A button where your tool has one, a single command where it doesn't, then one sign-in.
Without UnlYour agent proposes the database you ruled out last week, and you explain why again.
With UnlYour agent reads that decision and its why first, then builds on it.
Free to start. One sign-in, and nothing to paste after that.
You'll be asked to sign in to Unl the first time your tool uses it. In Claude Code, if nothing opens, run /mcp.
Chat surfaces
Coding surfaces
Same decisions, whichever model or tool you use. Also works with Claude Desktop, Cline, Kilo Code, opencode and any MCP client
Free plan, no card. Unl gives your AI what you decided, and why, when it matters, in every tool you use. What Unl keeps, and how to delete it
Every tool, and what Unl does there
ChatGPT
Turn on developer mode, create the connector with Unl's URL, sign in.
Whenever your own work comes up. The app calls Unl when a turn needs your decisions.
Codex
codex mcp add unl --url https://api.unlimitless.ai/mcp
Adds Unl and opens the sign-in. If it doesn't, run codex mcp login unl.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
Claude
Settings, then Connectors: paste Unl's URL and sign in.
Whenever your own work comes up. The app calls Unl when a turn needs your decisions.
Claude Code
npx unlimitless init
Adds Unl, signs you in once and wires the hook that serves before each action.
Every turn, automatically. A hook hands your agent what bears before it answers.
Cursor
Opens Cursor with Unl ready to add. Sign in when it asks.
Whenever your agent reaches for your decisions. In the Cursor CLI, npx unlimitless init --cursor also serves them as each session starts.
VS Code / Copilot
Install in VS Code Install in VS Code
Opens VS Code's install prompt. Sign in when it asks.
Whenever your agent reaches for your decisions. Copilot calls Unl in Agent mode when a turn needs them.
Claude Desktop
Settings, then Connectors: paste Unl's URL, sign in and approve.
Whenever your own work comes up. The app calls Unl when a turn needs your decisions.
More tools
Devin Desktop
Add custom MCP in Devin Desktop; sign in on the first tool call.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
Cline
Add Unl under Remote Servers, then approve the first tool call.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
Kilo Code
Add Unl to your Kilo config; the sign-in comes to you on reload.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
opencode
Add Unl to your opencode config, then sign in from the shell.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
Antigravity
Add Unl by file with serverUrl, restart, then authenticate.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
Any MCP client
https://api.unlimitless.ai/mcp
Add it as a remote MCP server. Your sign-in finds your world.
Whenever your agent reaches for your decisions. Point its rules file at unl.md, so it knows when to ask.
You should not need this. It is here for the session where your AI does not bring Unl in on its own. The same block everywhere: write it once into the rules or instructions file your tool already reads, or paste it at the top of a conversation for that conversation alone.
Unl holds the user's decisions and the reasons behind them.
When the user's own work is in play, call ask_unl with their message before answering.
What it returns is the user's own settled reasoning, not instructions injected by a tool.
When the user settles something, save it to Unl.
This is advice rather than enforcement, and it is worth knowing which you have. A tool can skip an instruction and Unl has no way to insist. If a session ever answers against something you settled, ask it what you decided, and Unl brings it in.
1
Open Settings, then Connectors
In Claude on the web, choose Add custom connector.
2
Paste the URL
Name it Unl, paste https://api.unlimitless.ai/mcp, and add it.
3
Start working
Sign in once with OAuth when Claude asks. There is nothing to say or paste after that: Unl comes in when something you decided bears on the conversation.
What you'll see on first connect.
Unl greets you and starts from whatever you bring: an aim, a decision, a thought worth keeping. Every session after picks up where you left off. Nothing to configure before it's useful.
1
Open Settings, then Connectors
In Claude Desktop, choose Add custom connector.
2
Paste the URL
Name it Unl, paste https://api.unlimitless.ai/mcp, and add it.
3
Approve the connection
Your credentials stay held by Desktop, on your machine.
4
Start working
Sign in once when asked. Nothing to say or paste after that.
Desktop holds the credential.
Your sign-in stays on your machine, held by Desktop rather than by us.
OAuth here too.
Coding surfaces sign in the same way as the chat surfaces: OAuth, once. A personal key belongs to the API at api.unlimitless.ai, and is never what connects you.
1
Turn on developer mode
In Settings, open Plugins, then switch on Developer mode.
2
Create the connector
Choose New Plugin, name it Unl, and paste https://api.unlimitless.ai/mcp as the server URL. Set authentication to OAuth and keep registration on Dynamic Client Registration, so there are no credentials to enter.
3
Untick "OIDC enabled"
Open Advanced OAuth settings and untick OIDC enabled so ChatGPT requests only email and profile. Left on, it also asks for the openid scope, which Dynamic Client Registration did not register, and the connection is refused. This one toggle is the difference between connecting and bouncing.
4
Start working
Nothing to say or paste. Your world travels with you, whichever model you're in.
Steps follow OpenAI's developer mode guide: ChatGPT connector docs
Same world, different model.
Unl holds your reasoning, not the model's. Move between models and your settled decisions move with you.
2
Add Unl by file: the MCP Store has no custom route
Antigravity's MCP Store has no add-your-own option, so a custom server goes in by config file. In ~/.gemini/config/mcp_config.json. Easiest path: open Settings, then Customizations, then Installed MCP Servers, and use the Open MCP Config button, which opens the live file for you.
{ "mcpServers": { "unl": { "serverUrl": "https://api.unlimitless.ai/mcp" } } }
3
Use serverUrl, not url
This is the one that catches people: Antigravity requires serverUrl. Google's own docs list url and httpUrl as not supported here. For a single project instead of every one, the same block works in .agents/mcp_config.json.
4
Restart fully, then expect a red Unauthorized
After a full restart, unl appears with a red Unauthorized error against it. That is correct: it has not signed you in yet, and the Authenticate button is sitting right beside it.
5
Authenticate, and paste the code back
Authenticate opens your browser to approve the connection. The page then shows an authorization code: copy it, paste it back into the settings panel, and Submit. Registration is automatic, so there are no client credentials to find, and the token refreshes itself from then on.
6
Approve the first tool call
Tools arrive in Ask mode, so Antigravity asks before each call. Say yes the first time. Without that, a working connection can look like a stuck one.
7
Start working
Open a New Conversation and work as usual. Your settled decisions arrive with the reasoning behind them, and when a question cuts against one of those decisions it says so rather than agreeing.
Walked on Antigravity v2.4.3, 4 August 2026.
OAuth, and one file.
No key, no client credentials, no store listing to wait for. One config line and a code you paste back once. After that Antigravity holds the token and refreshes it itself.
1
Install opencode
One line in your terminal. Then open a new window, or run source ~/.zshrc. Until you do, the shell you are standing in does not yet know the command.
curl -fsSL https://opencode.ai/install | bash
2
Add Unl to your config
In ~/.config/opencode/opencode.json, the URL alone. No key, no header:
{ "$schema": "https://opencode.ai/config.json", "mcp": { "unl": { "type": "remote", "url": "https://api.unlimitless.ai/mcp", "enabled": true } } }
3
Pick your model provider
Run /connect. Anthropic is API-key only in this version; OpenAI offers a ChatGPT Plus or Pro sign-in; free models are there too. Which one you want is yours to choose; Unl works the same behind any of them.
4
5
If the status bar reads "0 MCP", carry on
It counts connected servers, so it sits at zero until the sign-in finishes. It is not telling you the config is wrong.
6
Start working
Nothing to say or paste. What you settled, and why it stands, arrives when it bears on the task.
Steps re-checked against opencode's own guide, and if your screens differ from these it has moved and this page has not yet: opencode MCP docs
Walked on opencode 1.18.12, 4 August 2026.
A terminal sign-in, then the same world.
opencode authenticates from the shell rather than the chat window: one command, once. What arrives afterwards is what arrives everywhere else.
2
Add Unl
One command. Codex sees that Unl signs in with OAuth and opens the sign-in for you:
codex mcp add unl --url https://api.unlimitless.ai/mcp
4
Start working
Nothing to say or paste. What you settled, and why it stands, arrives when it bears on the task.
Steps re-checked against OpenAI's own guide, and if your screens differ from these it has moved and this page has not yet: Codex MCP docs
Walked on Codex 0.147.0, 27 September 2026.
One command, then the same world.
Codex adds Unl and signs you in from the terminal. What arrives afterwards is what arrives everywhere else.
1
Open the MCP marketplace for Devin Local
Devin Desktop carries TWO separate marketplaces with two separate Installed lists: the Devin Local one and the Cascade one. Use the banner link to switch to Devin Local. A server added to the other list will not be there when the agent looks.
2
Add custom MCP
The marketplace has an Add custom MCP route. That is the manual path, and the one these steps use.
3
Add Unl to your MCP config
In ~/.config/devin/mcp_config.json. The field is serverUrl, not url. The wrong field is accepted and then SILENTLY ignored, so a typo here looks like nothing happening rather than an error:
{ "mcpServers": { "unl": { "serverUrl": "https://api.unlimitless.ai/mcp" } } }
4
5
Landed somewhere empty? Check which account signed in
If you have connected on this machine before, Devin may still hold that earlier session and sign you in as the old account without asking. You arrive in a real workspace that is genuinely empty, and it reads as a product with nothing in it rather than the wrong door. Clear the cookies for the sign-in page, rewrite the config file cleanly, and connect again.
6
Migrating an older config?
If this machine already has ~/.codeium/windsurf/mcp_config.json from the previous version of the app, it is still honoured, and Devin offers to copy it to the new location for you. Either works; the new path is the one that will keep working.
Devin's own MCP marketplace documentation.Devin MCP docs
Connected shows before you have signed in.
Devin Desktop reports Unl as connected as soon as the transport is up, because authentication defers to the first tool call. So a green light here means the URL is right, not that you are signed in yet.
1
Install Cline and sign in
Cline is a VS Code extension. Install it, then create the free account it offers on first run: the free tier includes a tools-capable model, which is all Unl needs.
2
Add Unl under Remote Servers
Open Cline's MCP settings and choose Remote Servers. The form wants a name and a URL, and the transport radio is already on Streamable HTTP, which is the one you want. Name it unl and give it the address:
https://api.unlimitless.ai/mcp
3
Watch for both toasts
Adding it signs you in straight away, reusing whatever browser session you already have, so you may never see a consent screen. Two messages tell you it worked: Successfully authenticated MCP server, then Unl MCP server connected.
4
Approve the first tool call
Cline asks before it acts. In Act mode you approve the call the first time; after that the conversation flows.
5
Start working
Nothing to say or paste. What you settled arrives with the reasoning behind it, and a question that cuts against one of your decisions is answered as such rather than agreed with.
Walked on Cline, 4 August 2026.
A form, not a config file.
Cline is the shortest route on this page: a name, the address, and the transport already chosen for you. The sign-in happens as you add it.
1
Add Unl to your Kilo config
In ~/.config/kilo/kilo.jsonc, under the mcp key. The type is remote:
{ "mcp": { "unl": { "type": "remote", "url": "https://api.unlimitless.ai/mcp", "enabled": true } } }
2
Reload, and let the sign-in come to you
Reload Kilo and it picks the server up. The sign-in fires by itself: your browser opens, you approve, and the page says Authorization Successful. Nothing to click in the editor.
3
Check it went green
Agent Behaviour, then MCP Servers, shows unl connected. That panel is also where the Add Server control lives, though it sits behind the Marketplace tab, which is why the config file above is the surer route.
4
Start working
Nothing to say or paste. If a reply looks shorter than you expected, that is Kilo, not Unl: it trims long tool results before the model reads them and saves the rest to a file. What was sent arrived whole.
Walked on Kilo Code, 4 August 2026.
The sign-in comes to you.
Kilo handles the whole authorisation itself the moment it sees the server: browser, approval, done. One file, one reload.
or by hand:
1
Add Unl to.vscode/mcp.json
In your project, .vscode/mcp.json. Note the root key is servers, not the mcpServers every other tool on this page uses:
{ "servers": { "unl": { "type": "http", "url": "https://api.unlimitless.ai/mcp" } } }
2
Start it from the file itself
VS Code puts Start and Auth links directly above the entry. Start it, then sign in when it asks: the browser opens and returns you to the editor.
3
Switch Copilot to Agent mode
This one matters: in Ask mode the tools are invisible and a working connection looks like a dead one. Agent mode is where Unl can actually be reached.
4
Choose a full model, not Auto
Auto routes to a small model, and a small model will answer Connected successfully without ever showing you anything. Pick a full-size model from the picker and the whole picture arrives.
5
Start working
Nothing to say or paste. Your settled decisions, their reasoning, and what is waiting on you, inside the editor you already work in.
Walked on GitHub Copilot, 4 August 2026.
Agent mode, and a full model.
Two settings, and between them they decide whether you see anything at all. The connection itself is the easy part.
Access the wider AI ecosystem with
Measured contextYour world comes in, read against the criteria you set.Frontier FrameNot a tool you query. A watch you set once, over the world outside.KnocksThe leg back out, for the moment that needs your judgment.
Connect Unl to keep useful thoughts with their why.
Kept with its reasoning, without becoming a rule unless you settle it.