openclaw-cost-tracker-mcp

MCP server for AI agent token-cost telemetry across Anthropic, OpenAI, Gemini, Ollama, AWS Bedrock — per-agent attribution, per-provider breakdowns, spend-spike anomaly detection (per-agent median × threshold), cheaper-model routing recommendations with 30-day savings estimates, monthly forecast. Reads any cost-log JSONL with the standard {request_id, timestamp, provider, model, agent_id, prompt_tokens, completion_tokens, cost_usd} schema. OpenClaw operators get native ~/.openclaw/cost-logs/ parsing; other deployments wrap their provider calls with the included logging shim or commission a Custom MCP Build adapter. Keywords: Claude API cost, OpenAI cost tracking, AI agent FinOps, LLM cost attribution, token spend monitoring.

What it does

Production AI deployments rack up token cost across Anthropic + OpenAI + Gemini + Ollama, dozens of agents, hundreds of tool/skill calls. The bill arrives at month-end — and by then, the answer to "where did our money go?" is buried in 100k log lines. Provider dashboards show "$X spent on Anthropic" but not "which of our six agents drove 78% of that." This MCP server surfaces per-agent + per-provider cost attribution live, queryable in plain English from inside Claude or any MCP-aware client:

> claude: where did our LLM spend go this week?
[MCP tools: cost_overview + top_cost_drivers]

Total spend last 7 days: $42.18
By provider:
  Anthropic    $30.40 (72%) — claude-sonnet-4 dominates
  OpenAI       $9.20  (22%) — chat-bot agent
  Gemini       $1.86  (4%)  — cron-summarizer (cheap-route working)
  Ollama       $0.00  (local, free)

Top 3 cost drivers:
  data-extraction-agent      $28.50 (68%)
  chat-bot                   $9.20  (22%)
  cron-summarizer            $1.86  (4%)

1 anomaly flagged — see find_cost_anomalies for details.

> claude: any cheaper-routing opportunities?
[MCP tool: model_routing_recommendations]

Recommendation: data-extraction-agent currently runs claude-sonnet-4
with avg 400-token completions — extraction-style work that
gemini-2.5-flash usually handles at ~95% quality for ~5% the cost.
Estimated savings: $27.10/30d if migrated. Test on a 10% slice first.

Why openclaw-cost-tracker-mcp

Three things existing tools (provider billing dashboards, generic FinOps tools, custom scripts) don't do:

1. Per-agent attribution, not just per-provider totals. Provider dashboards show "$X spent on Anthropic" — they can't tell you which of your six agents drove 78% of that. Cost tracker reads OpenClaw's per-request cost-log JSONL and aggregates with the agent_id intact.

2. Cost-spike anomaly detection per agent. A single 120k-token paste into chat costs more than a week of normal traffic. The default 3x-median-per-agent threshold flags those before they show up in the month-end bill.

3. Routing recommendations grounded in actual usage. Generic "use cheaper models" advice is useless. This server identifies specific agents whose volume + completion-length pattern suggests a cheaper provider would deliver the same outcome, with concrete 30-day savings estimates.

Built for the production-AI operator running OpenClaw in production with real spend that matters.

Tool surface

Resources:

• cost://overview — 7-day snapshot
• cost://forecast — 30-day projection
• cost://anomalies — recent flagged anomalies

Prompts:

• diagnose-cost-spike — walk a recent spike to its root cause + corrective action
• weekly-cost-digest — 200-word weekly cost digest

Quickstart

Install

pip install openclaw-cost-tracker-mcp

Configure for Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "openclaw-cost": {
      "command": "python",
      "args": ["-m", "openclaw_cost_tracker_mcp"],
      "env": {
        "OPENCLAW_COST_BACKEND": "mock"
      }
    }
  }
}

Backends

JSONL log format

Each line is one JSON record:

{"request_id":"req-abc123","timestamp":"2026-05-04T12:34:56Z","provider":"anthropic","model":"claude-sonnet-4","agent_id":"data-extraction-agent","skill_id":"extract-structured-data","prompt_tokens":8500,"completion_tokens":600,"total_tokens":9100,"cost_usd":0.0345,"duration_ms":4823}

If your OpenClaw deployment doesn't emit this format, wrap your provider calls with a small logging shim — sample shim in examples/.

Roadmap

Need this adapted to your stack?

If your AI deployment doesn't use OpenClaw's cost-log format — different agent harness, custom logging, AWS Bedrock metering, vendor billing API — and you want the same attribution + anomaly + routing visibility, that's a Custom MCP Build engagement.

To engage:

1. Email temur@pixelette.tech with subject Custom MCP Build inquiry
2. Include: a 1-paragraph description of your stack + which tier you're considering
3. Reply within 2 business days with a 30-min discovery call slot

This server is part of a production-AI infrastructure MCP suite — companion to silentwatch-mcp (cron silent-failure detection) and openclaw-health-mcp (deployment health). Install all three for full operational visibility.

Production AI audits

If you're running production AI and want an outside practitioner to score readiness, find the failure patterns already present (cost overruns being one of the most common), and write the corrective-action plan:

Same email channel: temur@pixelette.tech with subject AI audit inquiry.

Contributing

PRs welcome. Backends are pluggable — see src/openclaw_cost_tracker_mcp/backends/ for the contract.

To add a new backend:

1. Subclass CostBackend in backends/<your_backend>.py
2. Implement get_entries()
3. Register in backends/__init__.py
4. Add tests in tests/test_backend_<your_backend>.py

Bug reports + feature requests: open a GitHub issue.

License

MIT — see LICENSE.

Related

• Production-AI MCP Suite (Gumroad bundle) — this server plus 4 others in one curated bundle with a decision tree, day-one drill, and Custom MCP Build CTA. $99, or $49 with LAUNCH50 for the first 30 days.
• silentwatch-mcp — cron silent-failure detection
• openclaw-health-mcp — deployment health
• openclaw-skill-vetter-mcp — ClawHub skill security vetting
• openclaw-upgrade-orchestrator-mcp — read-only upgrade advisor
• AI Production Discipline Framework — Notion template, $29 — the methodology these MCP tools implement (cost overruns are pattern P3.x in the catalog)
• SPEC.md — full server design

Built by Temur Khan — independent practitioner on production AI systems. Contact: temur@pixelette.tech