接入你的智能体
發布者自己的指令。不是 PlugMCP 端點。
用我們的金鑰與日誌託管一份副本。
npx -y @modelstat/[email protected]Claude Code 沒有本機指令的安裝連結。複製下方指令並在終端機中執行。
claude mcp add modelstat-mcp -- npx -y @modelstat/mcp@0.6.0在終端機中執行。
{ "mcpServers": { "modelstat-mcp": { "command": "npx", "args": [ "-y", "@modelstat/[email protected]" ] } }}claude_desktop_config.json
遠端 URL 還是本機指令？ 兩種安裝方式的差異
來自發布者
One command installs the daemon, pairs this machine, and wires the modelstat MCP into every AI tool you have (Claude Code, Claude Desktop, Cursor, Codex, …). Paste it into a terminal — or into Claude Code / Codex / Cursor and let it run.
curl -fsSL https://modelstat.ai/install.sh | sh
Windows (PowerShell): irm https://modelstat.ai/install.ps1 | iex
MCP server for modelstat — ask any MCP-compatible AI tool about your token spend.
The installer downloads a small static binary, verifies its SHA-256 checksum, pairs this machine, and installs a background service. No Node, no Python, no package manager — the binaries are self-contained. Re-run it any time to upgrade (auto-update is on by default).
That's it. Open your dashboard at modelstat.ai/dashboard. The daemon runs on macOS 13+, every glibc Linux distro, and Windows 10/11. Full options (modes, flags, headless installs): daemon/docs/INSTALL.md.
This is the public source for everything that runs on your machine:
modelstat daemon — the native Rust daemon: two static binaries (the collector and the summariser engine) that watch your AI-tool log files, price them, redact client-side, and upload metadata. The collector's modelstat mcp subcommand is also the local Model Context Protocol server, so Claude Desktop / Claude Code / Cursor / Cline / Continue / Zed can answer "how much did we spend on X?" in chat.@modelstat/sdk — the backend SDK: capture the LLM calls your own services make, redacted + compacted client-side before they leave the box.Why it's open. The code that reads your files should be auditable. The hosted service that aggregates your team's metadata is closed-source; everything that runs on your laptop is right here — read it, fork it, build your own binaries, or pin a commit and install from source. See LICENSE.
your AI coding tools on YOUR machine modelstat cloud
┌──────────────────────────┐ ┌─────────────────────────┐ ┌──────────────────────┐
│ Claude Code · Codex │ │ modelstat daemon │ │ analytics dashboard │
│ Cursor · Cline · Aider │ session │ • parse + price turns │ redacted │ spend & ROI grouped │
│ Windsurf · Zed · Copilot │ ───────▶ │ • redact (PII / keys) │ ───────▶ │ by activity · repo · │
│ Claude Desktop · … │ logs │ • summarise (see modes) │ HTTPS │ model · person — │
│ (logs already on disk) │ │ → tokens + abstract │ │ the charts above │
└──────────────────────────┘ └─────────────────────────┘ └──────────────────────┘
↑ secrets & PII are stripped on your machine before anything leaves ↑
Detects installed tools — scans ~/.claude, ~/.codex, ~/.cursor, ~/.aider, ~/.config/continue, and other tool-specific locations.
Parses local log files — Claude Code's JSONL logs, Cursor's SQLite DB, Codex's conversation files, and so on. The per-tool parsers are the only code that opens these files.
Redacts on-device — in every mode — every excerpt goes through a regex pass for secrets/PII and an on-device NER model before the uploader ever sees it. This never moves off your machine, regardless of the mode below.
Summarises — you choose where (at install) — the redaction in step 3 always runs locally; only where the summary is written differs:
Change it any time with modelstat mode <local|self-hosted|cloud>.
Uploads what the mode implies — per-session totals (tokens, model, cost, duration) + provenance always. In Local/Self-hosted, the redacted ≤240-char abstract. In Cloud, the redacted turns (so the server can summarise them). Never raw prompts, code, or secrets — redaction runs on-device first in every mode.
git clone https://github.com/modelstat/modelstat.git
cd modelstat/daemon && cargo build --release
Requirements: a stable Rust toolchain, and an account at modelstat.ai to pair with. (The SDKs build with their own toolchains — Node 20.18+/pnpm 10.7+ for sdks/node; Swift 5.9+ for the macOS tray.)
Full index: modelstat.ai/integrations
The installer wires the local MCP server (modelstat mcp) into every AI tool it detects — nothing to configure. No daemon on this machine? Connect the hosted server instead: add https://mcp.modelstat.ai to any MCP-compatible client and sign in via the browser on first use.
"How much did my team spend on Claude Code last week?" "Which project is driving my Cursor cost?" "Recommend a model for a code-review task — based on what worked for us before."
No in-band auth — the local server reuses the token the daemon already wrote on this machine. Details: mcp.modelstat.ai · source in daemon/crates/modelstat-mcp/.
You can verify, from this repository alone, exactly what does and doesn't leave your machine.
The boundary. Exactly one crate talks to our ingest endpoints: modelstat-ingest (DeviceApi::upload_batch). It POSTs to /v1/ingest in Local and Self-hosted modes (pre-summarised abstracts) or /v1/ingest/raw in Cloud mode (the redacted turns the server summarises). There is no other outbound data channel. The wire-format types live in daemon/crates/modelstat-wire/src/schema.rs — if a field isn't there, the uploader literally cannot send it.
What we receive — always: model/provider/tool, per-class token counts, the cost we computed, scrubbed cwd/git metadata, filenames (paths scrubbed), redaction counts (never the matched text), and a provenance stamp. Plus, depending on your mode: a ≤240-char on-device abstract (Local / Self-hosted), or the redacted conversation turns themselves (Cloud — that is exactly what our servers summarise for you).
What we never receive: your raw prompts, your code, or any API key / token / password. Redaction runs on your machine in every mode before anything leaves it — Cloud mode's uploaded turns are redacted turns, not raw ones, and Self-hosted mode runs the same on-device scrub before excerpts reach your org's endpoint. Text fields are size-capped by Zod; the paranoid policy drops entire stdout/stderr/tool_output/raw_text blobs before upload.
Defence-in-depth, every byte crossing the boundary passes through: parser scoping → secrets regex (Anthropic/OpenAI/Google/AWS/GitHub/Slack/Stripe keys, JWTs, PEM blocks, bearer headers, DB passwords) → PII regex (emails, public IPs, URL creds, home paths) → a Shannon-entropy catcher for unknown key shapes → on-device NER → wire-schema byte caps → provenance stamp. The regex + entropy floor lives in daemon/crates/modelstat-redact/; the caps in daemon/crates/modelstat-wire/src/caps.rs. In Cloud mode the daemon is fail-closed on the NER pass: if that on-device model can't load, it will NOT ship raw turns with regex-only redaction — it falls back to local extractive abstracts instead (no raw egress).
Audit it yourself — three places tell the whole story:
daemon/crates/modelstat-wire/src/schema.rs — every type that can be uploaded.daemon/crates/modelstat-redact/ — the redaction passes + patterns.daemon/crates/modelstat-ingest/ — the single upload path to our server.modelstat status prints, locally, the same token + redaction counters the server sees. Security disclosures: SECURITY.md · [email protected].
modelstat/
├── daemon/ The shipping daemon (Rust)
│ ├── crates/ collector CLI + daemon, summariser engine, MCP server,
│ │ parsers, redaction, wire schema, updater
│ ├── scripts/ install.sh / install.ps1 (served at modelstat.ai/install.{sh,ps1})
│ └── docs/ INSTALL.md and friends
├── apps/
│ ├── tray-mac/ macOS menu-bar app (Swift)
│ └── daemon/ retired TypeScript daemon (no longer shipped)
├── packages/ retired TypeScript packages — superseded by daemon/crates
├── sdks/ Backend SDKs (Node / Python / Rust) — current
├── prices/ Provider + model price tables
└── .github/workflows/ CI + daemon release builds + SDK publishing
See CONTRIBUTING.md. The short version:
[email protected] if sensitive).daemon/crates/modelstat-parsers/src/.modelstat is the cross-tool AI spend analytics layer for engineering teams — the questions your finance team actually asks: How much did we spend on project X? Which work-type eats our budget? Personal vs company accounts on the same machine? Opus vs GPT-5 vs Sonnet on the same kind of task?
MCP server for modelstat — ask any MCP-compatible AI tool about your token spend.
{ "mcpServers": { "modelstat-mcp": { "command": "npx", "args": [ "-y", "@modelstat/[email protected]" ] } }}工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
@modelstat/mcp 透過 stdio 在本機執行。
上次檢查時間：2026-11-14。 MCP server for modelstat — ask any MCP-compatible AI tool about your token spend.