fm-mcp

io.github.yranganav0.1.0更新於 Oct 10, 2026

Delegate summarising, extraction, classification and OCR to the macOS 27 on-device model.

概覽

AI 產生的概覽

讓程式開發代理把摘要、欄位擷取、分類與圖片 OCR 等簡單文字工作交給 macOS 27 上的裝置端 Apple Foundation Models 模型處理。

功能
這是一個本機 stdio MCP 伺服器,提供四個工具:summarise(把日誌、文件、筆記或逐字稿濃縮成一段文字或最多 7 個要點)、extract(依你提供的 JSON Schema 把具名欄位擷取成 JSON)、classify(把文字歸入你指定的標籤,可用 multi 選多個),以及 ocr(讀取 PNG、JPEG、HEIC、TIFF、GIF 或 BMP 圖片中的文字,不支援 PDF)。前三個工具可接收文字,也可接收最大 1 MB 的文字檔路徑,代理不必自行讀取檔案。首次呼叫工具時它會自行啟動 fm serve 模型服務,工作階段結束時停止。
適用情境
適合開發代理原本要花付費遠端 token 處理例行文字工作的情境,例如摘要日誌或從文件擷取欄位,同時希望文字留在本機。不適合程式碼、數學、推理、文字中不存在的事實,以及答錯代價高又無法核對的任務。
執行需求
需要 Apple Silicon 的 Mac(不支援 Intel)、已開啟 Apple Intelligence 的 macOS 27,並需先執行一次 fm license 接受授權。可透過 Homebrew、cargo 或安裝指令稿安裝;fm-mcp install 會設定 Claude Code 與 Codex,也可使用 Claude Code 外掛。選用環境變數 FM_MCP_REQUEST_TIMEOUT_SECS(預設 120)與 FM_MCP_LOG=debug。
安裝前請注意
裝置端模型較小且會出錯:摘要可能遺漏或扭曲細節,extract 對文字中只是意思接近的欄位可能給出錯誤值,重要結果請自行核對。輸入與回答共用約 8K token,同一時間只處理一個請求,安全過濾器有時會拒絕無害輸入。fm-mcp install 會寫入 ~/.claude.json、~/.claude/skills/、~/.codex/config.toml 與 ~/.codex/AGENTS.md,寫入前會備份;--dry-run 可預覽,--uninstall 可移除。文字僅透過本機 Unix 通訊端傳給本機模型。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 fm-mcp,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

fm-mcp

An MCP server that lets coding agents such as Claude Code and Codex hand simple text work to the on-device model for Apple Foundation Models, the fm that ships with macOS 27. It's free, private, and runs offline.

mcp-name: io.github.yrangana/fm-mcp

Why

Coding agents spend paid, remote tokens on simple jobs, like summarising a log or pulling fields out of a document. A small model already on your Mac can do much of that for nothing, and your text never leaves the machine. fm-mcp lets the agent hand off that work, and tells it plainly what the small model can't do.

Install

You need a Mac with Apple Silicon (Intel Macs are not supported), macOS 27 with Apple Intelligence turned on, and the fm licence accepted once (fm license).

sh
brew install yrangana/tap/fm-mcpfm-mcp install    # sets up Claude Code and Codexfm-mcp doctor     # checks everything fm-mcp needs

Then start a new Claude Code or Codex session. fm-mcp starts the model server (fm serve) itself the first time a tool is called, and stops it when the session ends.

Other ways to get the binary:

sh
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/yrangana/fm-mcp/releases/latest/download/fm-mcp-installer.sh | shcargo install fm-mcp

What fm-mcp install changes

  • Claude Code: adds the fm-mcp MCP server to ~/.claude.json (through claude mcp add when the claude command is there), and installs the fm-delegate skill in ~/.claude/skills/.
  • Codex: adds [mcp_servers.fm-mcp] to ~/.codex/config.toml, and a marked section in ~/.codex/AGENTS.md (or in AGENTS.override.md, if you use one).

It backs up each file before changing it and prints every path it touched; running it again changes nothing. --claude or --codex picks one agent, --dry-run shows the changes without writing, and --no-guidance skips the skill and the AGENTS.md section. fm-mcp install --uninstall removes everything it added.

The skill and the AGENTS.md section tell the agent when to delegate. Without them the agent sees only the tool descriptions, and delegates less well.

Or: the Claude Code plugin

The plugin gives Claude Code the MCP server and the skill, without fm-mcp install. It still needs the binary on your PATH, so install that first (Homebrew above); otherwise Claude Code reports the server as failed to connect.

/plugin marketplace add yrangana/fm-mcp/plugin install fm-mcp@fm-mcp

Use the plugin or fm-mcp install for Claude Code, not both: both together give Claude Code two fm-mcp servers. With the plugin, run fm-mcp install --codex to set up Codex only.

Other MCP clients

fm-mcp is a plain stdio MCP server, so other clients that run local servers on the same Mac can use it. fm-mcp install doesn't set them up, and they don't get the delegation guidance: the agent decides from the tool descriptions alone. Each client below was tried on a real Mac; others may work but are untested.

Claude Desktop (tried with 2.31226.1, 2026-10-11). In Settings › Developer › Edit Config, add this to mcpServers in claude_desktop_config.json, then quit and reopen Claude Desktop:

json
"fm-mcp": {  "command": "/opt/homebrew/bin/fm-mcp",  "args": []}

Use the full path (which fm-mcp prints it). fm-mcp helps most with files on your Mac: give Claude the file's path and it passes path, so the text never goes through Claude. For a web page, Claude Desktop reads the page itself first, so delegating it saves nothing.

Devin (formerly Windsurf) (tried with Devin.app 3.8.20, 2026-10-11). Add this to mcpServers in ~/.config/devin/mcp_config.json:

json
"fm-mcp": {  "type": "stdio",  "command": "/opt/homebrew/bin/fm-mcp",  "args": []}

Tools

ToolWhat it doesHow much it takes
summariseCondenses logs, documents, notes and transcripts into a gist: a paragraph, or at most 7 bullets. Not a full record.About 30,000 words of prose, but only about 1,000 lines of a dense log. Long input is split and combined, which takes a minute or two and can drop details.
extractPulls named fields out of one text as JSON, from a JSON Schema you give. Works best with a flat schema.About 3,500 words of prose.
classifySorts one text into labels you choose (one label, or several with multi).About 3,500 words of prose.
ocrReads the text in a PNG, JPEG, HEIC, TIFF, GIF or BMP image. Not PDF.One image.

summarise, extract and classify take text, or a path to a text file (up to 1 MB) so the agent never has to read the file itself.

Limits

  • Small model, small context: about 8K tokens, shared by the input and the answer. Logs full of IDs and numbers use far more tokens per word than prose.
  • Not for code, maths, reasoning, facts the text doesn't contain, or anything where a wrong answer is costly and you can't check it.
  • It makes mistakes. Summaries can drop or distort details. extract usually returns null for a field the text doesn't contain, but when the text has something close in meaning (no PO number, but line items), about 1 in 3 such fields gets a wrong value. Nested schemas are less reliable than flat ones. Check what matters.
  • One request at a time. The model is shared by every session on the Mac, so calls queue: a short extract takes 1–2 s, a long summarise a minute or more.
  • Safety filter: the on-device model sometimes refuses harmless input. fm-mcp reports that clearly, and the agent does the task itself.

Privacy

Everything runs on your Mac. fm-mcp talks to fm serve over a local Unix socket and sends nothing over the network. Your text goes only to the on-device model.

Troubleshooting

Run fm-mcp doctor. It checks the Mac, macOS, fm, the licence, the model, a test request, and both agents' setup, and says how to fix anything that fails.

ProblemFix
Licence not agreedRun fm license once.
Model not availableTurn on Apple Intelligence in System Settings and wait for the model to finish downloading.
A tool says the model "got stuck"Usually a schema with fields the text doesn't contain, or another session using the model. Try a flatter schema.
A tool times out on long inputRaise the limit: set FM_MCP_REQUEST_TIMEOUT_SECS (default 120) in the MCP server's environment.
You need more detailSet FM_MCP_LOG=debug in the MCP server's environment. fm-mcp logs to stderr only (stdout carries the MCP protocol).

Uninstall

sh
fm-mcp install --uninstallbrew uninstall fm-mcp

Development

sh
git clone https://github.com/yrangana/fm-mcp.gitcd fm-mcpcargo build --release./target/release/fm-mcp install    # points your agents at this build
cargo test --features fake-fm      # needs no Apple Intelligence: uses a fake fmcargo clippy --all-targets --all-features -- -D warningscargo fmt --checkscripts/real_fm_smoke.sh           # checks against the real fm on this Mac

AGENTS.md has the architecture, the tested facts about fm serve, and the project rules. docs/MANUAL_TEST.md is the checklist for each release.

Licence

MIT. fm-mcp is not affiliated with or endorsed by Apple. It calls the fm tool installed on your Mac and includes nothing from Apple.

來源:README.md,提交 831d5dd

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v0.1.0最新Oct 10, 2026