fuli MCP server

Each task gets only the small slice it needs. Your judgment and experience keep adding up, so FULI knows you better the more you use it. That is the compound interest. Little by little, what your AI makes starts to carry your taste.

What FULI does

Four FLAs come built in:

How it works

flowchart LR
    U[You] -->|ask inside a project| A[Agent client]
    A -->|task starts| F[FULI on your computer]
    F -->|preferences and a small recall| A
    A -->|look up more if needed| F
    A -->|write back what is reusable| F

A task starts with the preferences in effect and a few related facts. The Agent looks up the rest when it needs it, and at the end writes back only what the next task can reuse.

Install

FULI runs a small service and a console on your own computer.

Desktop client (recommended). Download from the latest release:

System	File
Windows	FULI-Setup-<version>.exe. No terminal and no administrator rights.
macOS (Apple silicon, 13.5 or later)	fuli-macos-arm64-<version>.dmg. Drag FULI onto Applications.

On first launch FULI finds the AI clients on this computer and connects each with one click. The installers are not signed yet, so each system holds FULI back once; Install and run shows the two clicks that let it through.

Servers, VMs and the cloud. Without a desktop, and on Intel Macs and Linux, install with npm (Node.js 24.12 or later):

npm install --global fuli-context
fuli setup
fuli open

fuli setup shows its plan and changes nothing until you confirm.

The console

The console opens on what matters now: questions waiting for you, the FLAs at work and what they finished. Everything runs on 127.0.0.1; there is no account and no sign-in.

FAQ

On your computer. Your personal graph and conversation records stay local. External knowledge (Notion, Lark, MCP or RAG sources) is read-only: FULI never writes back.

Claude Code, Codex, OpenCode, Cursor, DeepSeek Harness, Claude Desktop, Gemini CLI, Qwen Code, Kiro and Antigravity connect from the console (Settings › Clients) with one click. Other MCP clients run fuli connect for a standard configuration. What each client supports, such as capturing conversations, is in Connect a client.

Start a message with #FLA and its name. In Claude Code, OpenCode and Cursor you can also type /fla and pick one from the menu; in the Codex app, type /prompts:fla. Turn the menu off in Settings › Clients.

A session starts with only the preferences in effect and a few related facts; the Agent looks up the rest when it needs it. Facts that only you know, such as a sandbox URL you mentioned once, come back later without a search through the code. The numbers are in the benchmark (section 8.9.4).

By default the built-in FLAs work inside your client's session. Only when you connect a model for them do they use their own model, so nothing spends money behind your back.

No. Without one, FULI finds knowledge by its words, retrieval terms and relations. To find it by meaning, open Models and usage and choose Built-in models on Semantic search: FULI downloads a small model (GTE Small 中文 is recommended) and runs it on this computer. It needs no endpoint and no key, and your knowledge stays on the computer.

On the project's page, click Organise and pick one of your connected AI clients. Astro reads the project with that client's own model and fills in what is missing. You can add where the existing material is: folders, files or links. When several projects share a preference, it asks whether that preference should apply everywhere.

In the console, open Settings › Clients and choose Remove for that client. fuli uninstall and the Windows uninstaller do the same for every connected client. FULI takes out only its own entries (the MCP server, hooks, Skills and its permission lines), backs up each file first, and keeps a Skill you edited yourself. Your FULI data stays until you delete it. On macOS, remove the clients in the console before you move FULI.app to the Trash: the Trash alone leaves those entries in place.

The program is, your data is not: both use the same data folder. When the desktop client starts, it runs the one FULI service and points your connected clients at itself. The console then offers to remove the npm copy with one click; until it is removed, a bare fl in a terminal is still the npm version. Details

You do: AI looks things up and does the work, and humans retain final authority. Each task ends with one check that keeps only reusable knowledge. A preference an Agent guessed counts once you, or Tonborg within the autonomy you set, confirm it. Older answers stay in the history with their source and reason.

Learn more

Topic	Read
Install details, CLI commands, ports	Install and run
Principles and what is stored where	How FULI works
Connecting each client	Connect a client
Skills, hooks and MCP tools	Agent integrations
Notion, Lark and other knowledge bases	External knowledge
Tonborg and autonomy	Judgment and autonomy
The roundtable	Roundtable
What is proven so far	Capabilities and limits · Measured report
Running the source	Develop and verify · Contributing

The FULI mark and wordmarks come from the FULI 复利 brand system; see brand/README.md. Licensed under Apache-2.0.