modelstat

Know exactly what your AI coding spend bought. modelstat turns the session logs Claude Code, Codex, and Cursor already write into dollar-precise spend & ROI — broken down by the real work it went to, by project, and by model.

Local-first by construction. A small model on your machine summarizes and redacts every session before anything is uploaded. Raw prompts, code, and secrets never leave the box — only token counts, cost, and a short scrubbed abstract. The source is auditable on GitHub.

How it works

    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    │ • summarize (local LLM) │   HTTPS   │ model · person —     │
│ (logs already on disk)   │           │ → tokens + abstract     │           │ the charts above     │
└──────────────────────────┘           └─────────────────────────┘           └──────────────────────┘

                      ↑ raw prompts, code & secrets never leave your machine ↑

Install

One command installs the daemon, pairs this machine, and wires the modelstat MCP into every AI tool you have. Paste it into a terminal — or into Claude Code / Codex / Cursor and let it run.

curl -fsSL https://modelstat.ai/install.sh | sh        # macOS / Linux
irm https://modelstat.ai/install.ps1 | iex             # Windows (PowerShell) — MCP only

Already have Node? The package runner does the same — pairs the daemon and wires the MCP:

npx modelstat@latest        # or: bunx modelstat@latest  ·  pnpm dlx modelstat@latest

The only thing the curl script adds is bootstrapping Node when it's missing; with a runner you bring your own. Skip the MCP wiring with MODELSTAT_NO_WIRE=1 (or run it standalone any time: npx -y @modelstat/mcp wire).

The first run downloads the on-device summariser model (~2.7 GB Qwen GGUF to ~/.modelstat/models/), pairs the device, and installs a launchd user daemon on macOS (~/Library/LaunchAgents/ai.modelstat.daemon.plist) or a systemd user unit on Linux (~/.config/systemd/user/modelstat.service). It starts on login and watches your AI-tool logs in the background; the CLI then exits — there's no foreground process to keep open.

Requires Node 20+. macOS and Linux (x86_64, arm64). Then open modelstat.ai/dashboard.

Works with your real sessions

Claude Code · Codex · Cursor · Cline · Continue · Aider · Windsurf · Zed · GitHub Copilot · Claude Desktop. Nothing to instrument and nothing to intercept — modelstat reads the logs these tools already write. Per-tool setup: modelstat.ai/integrations.

Commands

npx modelstat@latest                     # install/upgrade + wire the MCP. Default action.
npx modelstat@latest remove              # stop and uninstall the background service

npx modelstat@latest status              # pairing, service + live usage: sessions · tokens · cost
npx modelstat@latest jobs                # pipeline queue + recent processing ledger
npx modelstat@latest paths [--json]      # state file + log dir + API URL

npx modelstat@latest sync --session <id> # force-ingest ONE session now (warms a running daemon)
npx modelstat@latest reset               # reset cursors so the daemon re-reads everything
npx modelstat@latest watch               # foreground watcher (no service install)
npx modelstat@latest discover            # report detected tool installs + identities
modelstat statusline                     # Claude Code status line (reads its stdin JSON)

Headless pairing: npx modelstat@latest --json --no-browser emits one NDJSON event per line so a wrapper can drive pairing non-interactively.

Claude Code status line

The installer auto-enables a live status line in Claude Code so every turn shows your current session's tokens · effective $ · taxonomy. It reads only a small local cache (~/.modelstat/sessions/<id>.json) — never blocks the prompt, never calls the network. It composes with (and restores) any status line you already had:

{ "statusLine": { "type": "command", "command": "modelstat statusline" } }

Opt out at install with MODELSTAT_NO_STATUSLINE=1, or remove it later with npx modelstat@latest remove.

MCP — ask any AI client about your spend

Pair the daemon, then add @modelstat/mcp to query your own spend from inside Claude Code, Claude Desktop, Cursor, Cline, Continue, or Zed:

claude mcp add modelstat -- npx -y @modelstat/mcp

Full per-client docs: mcp.modelstat.ai.

Shared state across install methods

All daemon state lives in one home directory, ~/.modelstat, identical on every OS. identity.json holds your device UUID + bearer token (0600); state.json holds runtime state (file cursors, …). Set MODELSTAT_HOME to relocate everything at once (e.g. MODELSTAT_HOME=/opt/modelstat for a system-wide, one-per-server install). Installing via both Homebrew and npm yields the same binary reading the same state, and the service deduplicates the device server-side — you won't see the same laptop twice.

Self-host

Point the daemon at your own modelstat API instead of the hosted service:

export DAEMON_API_URL=https://your-modelstat-api.example.com
npx modelstat@latest

DAEMON_API_URL can also be set persistently via .env or in the systemd/launchd unit.

Privacy

• Reads local session logs the tools already write — nothing is intercepted.
• Uploaded: token counts, model name, timestamps, provider-assigned session id, redactable git remote, and a redacted work-type abstract.
• Never uploaded: prompt text, model responses, file contents, tool-call arguments, environment variables, SSH keys, secrets.
• Redaction runs on-device; offline sessions buffer locally and upload when the network returns.

License

Apache-2.0. Source at https://github.com/modelstat/modelstat/tree/main/apps/daemon.