repofit MCPサーバー

Stop scrolling trending. Ask the agent that has watched you work.

Every week another wave of repositories hits ten thousand stars, and every "tools you should try" list is written for everyone at once. Meanwhile your coding agents have been taking notes on you for months: what you build, what keeps breaking, what you complained about at 2 a.m. repofit reads those notes on your own machine, from Claude Code, Codex, Cursor and a dozen more, and turns them into picks you have not heard of and will actually use. Every pick shows you the moment in your own history that earned it.

Try it in one line

uvx repofit unused

No install, no account, no API key and no model. It reads the session logs your agents already keep and lists the MCP servers and skills you installed but never called. Each one still costs context on every request. Without uv, run pipx run repofit unused.

Read 17 sessions from Claude Code, Codex, Gemini CLI. Nothing left this machine.

7 of 9 installed MCP servers and skills were never called:

  Claude Code · 12 sessions
    MCP server   sqlite
    skills       canvas-design, docx, pptx
  Codex · 4 sessions
    MCP server   puppeteer
    skills       slack-gif-creator, xlsx

Each one still costs context on every request. Remove the ones you do not need.

Not checked: 1 tool in Gemini CLI. Its logs do not say which server or skill each call went
to, or hold no sessions yet.

Want GitHub projects picked from the same history? Connect `repofit mcp` to your agent:
https://github.com/pth2002/repofit

Output on a sample history.

https://github.com/user-attachments/assets/c7ac0bcf-74ed-4a09-9295-b6d06e6d67c1

Install

pipx install repofit

Then point the agents you use at repofit mcp. Most of them read the same MCP config:

{ "mcpServers": { "repofit": { "command": "repofit", "args": ["mcp"] } } }

That line goes in ~/.cursor/mcp.json for Cursor, ~/.gemini/settings.json for Gemini CLI, ~/.qwen/settings.json for Qwen Code, and the MCP settings of Cline, Windsurf or Claude Desktop. Codex reads TOML, and Claude Code adds servers from the command line:

# ~/.codex/config.toml
[mcp_servers.repofit]
command = "repofit"
args = ["mcp"]

claude mcp add repofit -- repofit mcp

repofit runs inside your agent, so your agent's own model does the thinking: no API key, no account, nothing to host. To skip the install step, use uv and make the command uvx repofit mcp.

Just ask

Open a new session in any agent you connected and talk to it like a colleague who has read your history:

• "Use repofit to recommend some GitHub projects for me."
• "Does github.com/owner/name fit me?"
• "Is there a good tool for turning PDFs into Markdown?"
• "Which of my installed skills do I never use?"

The agent shows you the picks in the chat, and the same list stays one command away:

The picks come from a sample history. The text is exactly what the CLI prints.

Every pick has a receipt. No "because you starred X": the seen in line points at the session where the need showed up and quotes what you said. If repofit cannot point at your history, it does not make the pick.

The list at the bottom matters more than it sounds. Every skill and MCP server you installed and forgot still costs context on every single request, and repofit names the ones none of your sessions ever called.

How it picks

Each agent stores its history differently, and the formats change every few months, so repofit ships one adapter per agent. The profile is written from a redacted, numbered evidence pack, and any claim that cites evidence that does not exist is thrown out by the code, not left to the model's good manners. Retrieval runs one query per task you actually do, with project names and search phrases the model adds, so it finds the library a task depends on even when you never typed its name. The ranker is told to skip the household names you already know. The candidate pool is rebuilt every Monday from the GitHub API: active repositories above 3,000 stars, new ones above 1,000, the agent ecosystem above 300, plus GitHub trending.

Your history stays yours

• Everything runs on your machine. repofit reads the files your agents already keep and writes its own to ~/.repofit.
• You choose the depth: installed tools only, plus memory and instruction files (the default), or plus your own messages and tool calls.
• Before any model sees a line, API keys, tokens, passwords, emails, phone numbers and random-looking strings (yes, including the password you once pasted after ssh) are stripped, and your home folder becomes ~.
• With MCP, the text goes to the model your agent already uses. From the command line it goes only to the provider you name, after you confirm.
• Your notes and other people's READMEs can hide instructions. repofit hands everything over as data, and the agent is told never to follow what it finds inside.

Supported agents

Agent	Sessions	Memory and instructions	Installed tools
Claude Code	yes	CLAUDE.md, project memory	skills, plugins, MCP servers
Codex	yes, CLI and desktop	AGENTS.md, memories/	skills, plugins, MCP servers
Cursor	yes	user rules, pending memories	MCP servers, extensions
Hermes Agent	yes	USER.md, MEMORY.md	skills, MCP servers
Gemini CLI, Qwen Code	yes	GEMINI.md, QWEN.md	MCP servers, extensions
Cline, Roo Code, Kilo Code	yes	.clinerules	MCP servers
Continue	yes	rules/	MCP servers
VS Code Copilot	yes	copilot-instructions.md	MCP servers, extensions
Kimi CLI, Aider	yes
OpenCode	counted	AGENTS.md	MCP servers
Windsurf, Claude Desktop, WorkBuddy, Trae	no local sessions	memories where present	MCP servers

Project instruction files (AGENTS.md, CLAUDE.md, .cursorrules and friends) are read too. Run repofit scan to see what it finds on your machine. If your agent hides its data somewhere else, open an issue with the path and it gets an adapter.

Install with pipx install repofit, or put uvx in front of each command.

repofit unused                                  # installed MCP servers and skills no session ever called
repofit scan                                    # which agents are installed and what each one stores
repofit profile build --level L2 --provider deepseek
repofit recommend --provider deepseek --lang zh

Providers: deepseek, openai, ollama, anthropic, or any OpenAI-compatible endpoint with --provider compatible --base-url ... --model .... Keys come from DEEPSEEK_API_KEY, OPENAI_API_KEY or ANTHROPIC_API_KEY, and repofit shows how much text it will send before it sends anything.

No provider at all? Each model step prints a prompt to paste into any chat and imports the answer:

repofit profile prompt --level L2 > prompt.md   # then: repofit profile apply answer.json
repofit profile expand --prompt                 # then: repofit profile expand --apply answer.json
repofit recommend --prompt --lang en            # then: repofit recommend --apply answer.json
repofit recommend --no-llm                      # keyword retrieval only, nothing leaves your machine

Development

git clone https://github.com/pth2002/repofit && cd repofit
python -m pip install -e . pytest
python -m pytest
repofit index build      # rebuild the candidate pool yourself; needs a GitHub token

License

MIT