Unl

Name: Unl
Author: unlimitless

@unlimitlessMCP server · External

Your why agent, the infra agent for your fleet. Your agents work from what you decided, and why.

Product managementv0.2.3MITJavaScript

Freeno charge per call

One address for any MCP client. You sign in at clerk.unlimitless.ai once, after adding it.

The two-minute demo running in a terminal — From the README

What it does

Reads
Writes
High impact

At most: only reads or looks up information.

What it costs

Free

Odel charges nothing per call. The publisher may charge for its own service.

What it asks for at sign-in

Access to	Read	Change
Email address
Profile

From the sign-in at clerk.unlimitless.ai, as the endpoint publishes it. You approve it there.

Tools

None listed · checked 11 Oct 2026

Odel could not list its tools without signing in.

Try it in the Inspector

README5 min readView on GitHub

Unl, your why agent. Your AI already knows how to reason. Unl gives it your reason.

See it work in two minutes, no account: node demo/changed-decision.mjs

The two-minute demo running in a terminal

You decide things while you work with AI: what to build, what to leave out, and why. Unl keeps those decisions with their reasons. When you or one of your agents is about to act, Unl hands over the ones that bear on the task, so Claude Code, Cursor, ChatGPT, Codex and the agents you build all start from the same why. You never re-brief them.

What is open, and what is hosted

Open, in this repository, under MIT: everything that runs on your side. The unl command, the AI SDK package, the Cursor plugin, a working example for each agent framework, and the decision format with a conformance check you can run on your own code. Read them to see exactly what leaves your machine.

Also open: the status checks, which test Unl from outside its own infrastructure.

Hosted: Unl itself, the service that keeps your decisions with their reasons and chooses which ones bear on a task, runs at api.unlimitless.ai. Its code is not in this repository.

Start

With the agent you already use, tell it:

text

Read unlimitless.ai/start.md, set up Unl, and show me the decisions you find.

In Claude Code or Cursor:

bash

npm install -g unlimitless
unl              # sign in once, then the Unl terminal opens
unl init         # in your project, for Claude Code (add --cursor for Cursor)

Then ask your agent what you are working on.

In any MCP client (ChatGPT, Claude, Codex, VS Code, Grok Bot, Lovable): add the server https://api.unlimitless.ai/mcp and sign in when it asks.

In an agent you build: make a key at unlimitless.ai/portal/keys, set it as UNL_KEY, and pick your framework below.

Connect your tools

Every tool uses the same address, https://api.unlimitless.ai/mcp, and signs in with OAuth once. There is no key to paste: a personal key is only for the API, never for connecting a tool.

Claude and Claude Desktop. Open Settings, then Connectors, and choose Add custom connector. Name it Unl, paste https://api.unlimitless.ai/mcp, and add it.

Claude Code. In your project, unl init signs you in and wires the hook that brings Unl into every turn. To add only the server:

bash

claude mcp add --transport http unl https://api.unlimitless.ai/mcp

then run /mcp, choose unl and pick Authenticate.

ChatGPT. In Settings, open Plugins and switch on Developer mode. Choose New Plugin, name it Unl, paste https://api.unlimitless.ai/mcp as the server URL, set authentication to OAuth and keep Dynamic Client Registration. Then open Advanced OAuth settings and untick OIDC enabled: left on, ChatGPT asks for a scope Unl does not register and the connection is refused.

Cursor. Add the URL to .cursor/mcp.json, no key and no header, or install the Cursor plugin in this repository:

json

{ "mcpServers": { "unl": { "url": "https://api.unlimitless.ai/mcp" } } }

Everything else (Codex, VS Code, Antigravity, Grok Bot, Lovable, Viktor and more): step-by-step for each, kept current, at unlimitless.ai/connect.

What is here

Folder	What it is
`cli/`	The `unl` command. `unl init` adds Unl to a Claude Code or Cursor project, and `unl` on its own opens the Unl terminal: each turn, the decisions Unl handed your agent with their reasons, and the ones it held back.
`packages/ai-sdk/`	`@unlimitless/ai-sdk`. Wrap any Vercel AI SDK model with `withUnl` and every call carries the part of your decisions that bears on it.
`examples/`	One small working example per framework: Mastra, OpenAI Agents SDK, Google ADK, LangChain, Pydantic AI, LiteLLM, Cloudflare Agents, Netlify, TanStack AI, Composio, E2B, Browser Use and a GitHub Copilot custom agent.
`examples/vercel-ai-sdk-chat/`	A chat app on the AI SDK you can deploy to Vercel with Unl already added.
`.cursor-plugin/` and `mcp.json`	The Cursor plugin: Unl's MCP server, with sign-in discovered from the server. No key in the repository.
`decision-format/`	The open format for a kept decision: what it holds, how its address is computed and checked, a conformance check you can run on your own implementation, and reference code in Node and Python. MIT, and it runs without Unl.

Try the client contract locally

bash

git clone https://github.com/Unlimitless-Inc/unl.git
cd unl/packages/ai-sdk
npm install
npm test

These tests use synthetic responses. They test the client contract without an account or a model call, and do not prove hosted retrieval or model performance. The planner and writer example is the next step for trying a real changed decision in your own test project.

For example, you first choose synchronous exports because the files are small. Later you choose a queue because larger files exceed the request timeout. The job is for a fresh writer to receive the new choice and its reason, then produce a change that follows it. Compare with an equally updated file as well as stale or absent context. A maintained file may work just as well.

How an agent gets your why

Two ways in, and most examples show both:

On every call. A middleware, hook or instructions function asks Unl what bears on the turn and puts the answer in front of the model before it acts.
Through Unl's tools over MCP. The agent can open the full reasoning behind a decision, and propose a new one for you to confirm. Nothing becomes one of your decisions until you say so.