modelstat

See every AI token your team spends. Local companion for modelstat — reads the session logs your AI coding tools already write (Claude Code, Codex, Cursor, Cline, Continue, Aider, Windsurf, Zed, Copilot, Claude Desktop), tokenises events on-device, and uploads only metadata to your modelstat dashboard.

Your prompts never leave your machine. The agent uploads only token counts, model ids, timestamps, and a provider-assigned session id. Source is auditable on GitHub.

Install

The fastest way:

curl -fsSL https://install.modelstat.ai | sh

Or via Homebrew:

brew install modelstat/tap/modelstat
modelstat connect

Or via npm / pnpm / bun:

npm install -g @modelstat/agent && modelstat connect
pnpm add -g @modelstat/agent && modelstat connect
bun add -g @modelstat/agent && modelstat connect

modelstat connect installs a launchd user agent on macOS (at ~/Library/LaunchAgents/ai.modelstat.agent.plist) or a systemd user unit on Linux (at ~/.config/systemd/user/modelstat.service). The daemon starts automatically on login.

Requires Node 20+ (or skip Node entirely with the Homebrew route). macOS and Linux (x86_64, arm64) supported.

Commands

modelstat connect              # pair + install background service, then exit
modelstat status               # show pairing + service state
modelstat paths [--json]       # print state file + log paths + API URL
modelstat stop                 # stop and uninstall the service
modelstat scan                 # one-shot parse + upload of local JSONL
modelstat watch                # foreground watcher (no service install)
modelstat discover             # report detected tool installs + identities

Programmatic pairing (used by harness skills for OpenClaw, NanoClaw, etc.):

modelstat connect --json --no-browser

Emits one NDJSON event per line. Schema documented at integrations/harness-skills/modelstat-connect/README.md.

Shared state across install methods

Installing via both Homebrew and npm on the same laptop produces the same binary reading the same state file — on macOS that's ~/Library/Preferences/modelstat-agent-dev-nodejs/config.json. Your device UUID, bearer token, and pairing state persist across install methods. No duplicate devices on the server side either — the API upserts on (user_id, machine_id).

MCP server

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

# Claude Code
claude mcp add modelstat -- npx -y @modelstat/mcp

Full wire-up docs per client: https://modelstat.ai/mcp

Self-host

To point the agent at your own modelstat API (not the hosted SaaS):

export AGENT_API_URL=https://your-modelstat-api.example.com
modelstat connect

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

Privacy

• Agent reads local session logs written by the tools you use. Nothing is intercepted — the tools already write these files.
• Upload payload: token counts, model name, timestamps, provider-assigned session id, git remote URL (redactable), redacted work-type summary.
• Never uploaded: prompt text, model responses, file contents, tool-call arguments, environment variables, SSH keys, secrets.
• Redaction is on-device via @modelstat/parsers.
• Offline mode: buffered locally, uploaded when network returns. modelstat scan --dry-run for local-only analytics.

Pricing

• Free: 100M tokens/month, 1 device, no card.
• Team: $5/seat/month with 250M pooled tokens; overage at $25/billion.
• Enterprise: SSO/SCIM, on-prem ingest, SLAs — contact [email protected].

License

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