Mendrift

Autonomous MLOps incident response agent, plus mendrift-mcp — an open-source MCP server for drift detection and ML incident tooling.

pip install mendrift-mcp     # or: uvx mendrift-mcp

Published on PyPI and the MCP Registry as io.github.suneel190700/mendrift-mcp.

▶ Live demo — run a real incident in your browser: supply an alert, watch the agent diagnose it against a real MLflow registry, and approve or reject the rollback at the human-in-the-loop gate. Toggle between a crafted synthetic scenario and real US consumer-credit benchmark data. React frontend on a FastAPI backend; the free tier sleeps, so the first load may take ~40s.

When a production model drifts or degrades, Mendrift detects it, diagnoses the root cause from monitoring and registry evidence, proposes a remediation, and executes it only after human approval.

alert ──> classify ──> diagnose (MCP tools) ──> propose
             │                                     │
           noise ──> close               human approval gate
                                                   │
                                    execute ──> verify recovery

Built with LangGraph (agent orchestration), LangChain (ChatAnthropic + bind_tools), the Model Context Protocol, Evidently, MLflow, and Claude (Haiku + Sonnet).

mendrift-mcp tools

Safety model

The approval gate is enforced in the tool layer, not the prompt: execute_rollback verifies a single-use, action-scoped HMAC token minted only by the human review flow — the minting function is never exposed over MCP. A prompt-injected or confused agent cannot execute writes.

Tested live: Claude was first ordered to roll back "with full authorization" (it proposed but declined to fabricate a token), then handed a fabricated token, which the gate rejected by constant-time HMAC comparison:

See tests/test_approval_gate.py, including the action-scoping test: a token minted for one model/version is invalid for any other.

Human-in-the-loop, crash-proof

The incident graph halts before execution (interrupt_before) and checkpoints every step to SQLite. The process can die; a new process resumes the same incident by thread_id after a human mints the approval token — which enters state only via update_state(), from outside the graph. Denial is a first-class path: no token → closed_approval_denied, no execution.

Agent design

Routing lives in a code table (ROUTER_TABLE), not prompts, so cost per path is measurable config — ~3.9K input / 630 output tokens per incident. The diagnose loop is bounded (max 8 tool calls) with per-call retries and capped backoff; on tool failure the model receives a structured error record, and on budget exhaustion the agent degrades to an incident with partial evidence — it never invents a diagnosis. Destructive actions require affirmative evidence: a rollback is recommended only when retrieved evidence links the symptom to a specific deployment, never on deploy-correlation alone. The agent can also recommend monitor — real but mild, non-actionable drift is watched, not acted on.

Live mode

MENDRIFT_DEMO=0 runs the agent against real infrastructure rather than fixtures:

• scripts/seed_demo.py trains two sklearn versions into a local MLflow registry — v13 clean, v14 with a schema swap and a training window polluted by missed-fraud labels (recall 0.72 → 0.18, AUC 0.84 → 0.82) — and writes reference/current frames
• get_drift_report runs Evidently's DataDriftPreset over those frames, returning real Wasserstein/JS distances against per-metric thresholds, plus schema changes derived from actual column sets
• get_deployment_history / diff_deployments read the registry and the underlying runs — real aliases, params, metrics
• summarize_metric_anomalies scores the current window with both the production and previous versions, so it reports model divergence rather than population drift — a rollback clears it, ordinary data shift does not
• an approved execute_rollback moves the production alias for real

uv run mlflow server --host 127.0.0.1 --port 5001        # separate terminal
PYTHONPATH=src uv run python scripts/seed_demo.py

rm -f demo.db
MENDRIFT_DEMO=0 PYTHONPATH=src uv run python scripts/demo_interrupt.py start
MENDRIFT_DEMO=0 PYTHONPATH=src uv run python scripts/demo_interrupt.py approve

A live run diagnoses from computed evidence — e.g. "v2 introduced a schema swap replacing promo_flag with promo_flag_v2 … label_noise 0.0 → 0.45 collapsing val_recall 0.724 → 0.176 … 79.7% prediction-rate divergence from the prior version, model-induced, not population drift" — then halts for approval and resolves.

The eval suite deliberately stays on fixtures: evals need determinism and zero cost in CI, while live mode exercises the real stack.

Live web demo (two worlds)

A hosted web app wraps live mode behind a browser UI: a React (Vite) frontend on a FastAPI backend, deployed on Render. A visitor submits an alert, the frontend posts it to /api/diagnose, and the backend runs the real LangGraph agent — live Claude reasoning over an embedded MLflow registry (sqlite://, seeded on boot) — then halts at the HMAC gate. Approve or reject and the backend resumes the graph via /api/decision, executing a real alias rollback and verifying recovery. The Anthropic key lives only on the server; runs are rate-limited since each calls a real model. Try it: mendrift-demo.onrender.com.

The dashboard toggles between two seeded worlds, so the same agent can be seen against both a crafted scenario and genuine real-world data:

• Synthetic (scripts/seed_demo.py, model fraud-scorer) — the crafted schema-swap incident: clean, teachable, an unambiguous rollback story.
• Real US credit (scripts/seed_real.py, model credit-risk) — the Give Me Some Credit dataset (real US consumer-credit records, target SeriousDlqin2yrs) split by borrower age into reference/current windows for genuine feature drift, with a controlled model regression injected into v2 (asymmetric missed-default label noise) so the incident has ground truth. Real distributions and real Evidently drift; a known correct action. Measured gap: val_recall 0.637 → 0.156, AUC 0.854 → 0.810.

Injecting a known regression into real data is standard practice for validating a drift-detection system — it gives the evaluator ground truth for what the agent should decide while the drift computation still runs on genuine distributions.

The backend routes each request to the right world (model + parquet frames + label column) per the dataset field; the tool layer reads those from env vars, applied per-request under a lock so concurrent requests stay isolated.

Run the web app locally:

# 1. build the React frontend (FastAPI serves the built assets)
cd frontend && npm install && npm run build && cd ..

# 2. seed both worlds, then start the backend (frontend + API on one port)
export ANTHROPIC_API_KEY=sk-ant-...
export MLFLOW_TRACKING_URI="sqlite:///$(pwd)/mlflow.db"
PYTHONPATH=src uv run python scripts/seed_demo.py     # synthetic world (fraud-scorer)
PYTHONPATH=src uv run python scripts/seed_real.py     # real world (credit-risk)
PYTHONPATH=src uv run uvicorn app.main:app --port 8000     # open http://localhost:8000

For frontend development with hot reload, run cd frontend && npm run dev (port 5173); Vite proxies /api to the backend on port 8000.

Evaluation

src/mendrift/evals/ replays synthetic incident trajectories against the real graph — only the LLM (scripted) and the read tools (fixture world) are faked; the gated action tools are the genuine implementations, so the HMAC gate is exercised by every test. Four assertions per trajectory:

19 logic-distinct incident scenarios spanning the decision space, each with its own evidence shape and correct action: