monk-ai 0.2.0

Find hidden cost leaks and blind spots in your agentic AI workflows

🕵️ monk

Find hidden cost leaks and blind spots in your agentic AI workflows.

Drop monk on any trace file. Get a plain-English report of what's wasting tokens — and exactly how to fix it.

$ monk run ./traces/

  🕵️  monk — Agentic Workflow Blind Spot Detector
  Source: ./traces/   |   Calls analysed: 2,847

  ┌─────────────────────────────────────────────────────────────────────────┐
  │  3 blind spots found  ·  ~$62.40/day estimated waste  ·  ~$1,872/month  │
  └─────────────────────────────────────────────────────────────────────────┘

  🔴 [1] Retry loop: 'web_search' called 4x in a row  ·  ~$38.20/day
  Fix: Add a max-retries guard before calling 'web_search'.

  🔴 [2] 'get_user_profile' returns empty 80% of the time  ·  ~$19.10/day
  Fix: Guard against empty returns — don't pass null context back to the LLM.

  🟡 [3] Model overkill: gpt-4o used for simple tasks (62% of calls)  ·  ~$5.10/day
  Fix: Route classify/format calls to gpt-4o-mini — identical quality, 16x cheaper.

---

What monk detects

Detector	What it finds
retry_loop	Same tool called 3+ times in a row — agent stuck, burning tokens
empty_return	Tool returns null/empty and agent retries anyway
model_overkill	Expensive model used for short, simple tasks
context_bloat	System prompt consuming >55% of token budget, or unbounded history growth
agent_loop	Agent cycling through the same step sequence without progress

---

Install

pip install monk-ai

Or run from source:

git clone https://github.com/Blueconomy/monk
cd monk
pip install -e .

---

Usage

Analyse a single file:

monk run agent_traces.jsonl

Analyse a folder of traces:

monk run ./traces/

Run specific detectors only:

monk run traces/ --detectors retry_loop,model_overkill

Export findings as JSON (great for CI):

monk run traces/ --json findings.json

Only show high-severity findings:

monk run traces/ --min-severity high

Use in CI — monk exits with code 1 if high-severity findings exist:

- name: Run monk
  run: monk run ./traces/ --min-severity high

---

Trace format

monk auto-detects OpenAI, Anthropic, and LangSmith trace formats.

For custom logging, any JSONL file with these fields works:

{"session_id": "abc123", "model": "gpt-4o", "input_tokens": 1200, "output_tokens": 80, "tool_name": "web_search", "tool_result": "some result"}

Supported fields:

Field	Required	Notes
model	✅	e.g. gpt-4o, claude-sonnet-4-6
input_tokens	✅	Prompt token count
output_tokens	✅	Completion token count
session_id	Recommended	Groups calls into sessions
tool_name	Optional	Name of tool called
tool_result	Optional	Result returned by tool
system_prompt_tokens	Optional	Enables context_bloat detection

monk also natively parses:

• OpenAI Chat Completions API response format
• Anthropic Messages API response format
• LangSmith run export format

---

Why monk?

Most observability tools show you dashboards and traces. monk finds the things you don't know to look for.

The patterns monk detects are the ones we found repeatedly when auditing real agentic workflows — retry loops that nobody noticed, tools returning empty results silently, expensive models doing simple jobs. These don't show up as errors. They just quietly burn money.

---

Roadmap

• Live mode: instrument running agents via OpenTelemetry
• Prompt compression suggestions
• Cross-workflow benchmarking
• Slack / PagerDuty alerts
• Web dashboard

---

Contributing

PRs welcome. See CONTRIBUTING.md.

To add a new detector:

1. Create monk/detectors/your_detector.py extending BaseDetector
2. Add it to monk/detectors/__init__.py
3. Add tests in tests/test_detectors.py

---

License

MIT — free to use, modify, and distribute.

---

Built by Blueconomy AI — Techstars '25.
If monk finds something useful, give us a ⭐ — it helps more people find it.