Mesedi

Guardians for autonomous AI agents.

Mesedi observes agent executions in production, detects when an agent is going wrong (loops, crashes, runaway costs, prompt-injection attacks, validator failures, drift, silent tool failures), clusters related failures so the same problem doesn't page you a hundred times, optionally halts a misbehaving execution before it burns more budget, and notifies you out-of-band via webhook the first time a new failure class appears.

The Mesedi v1 surface is "detect, cluster, optionally stop, escalate, drop-in adopt." Per-failure-class canonical fix descriptions (Tier 1 Playbooks) ship in the dashboard; auto-fix (Tier 3) is on the v2 roadmap.

Repository layout

This is a monorepo. Each top-level directory is independently buildable; cross-directory changes (e.g. backend wire-format updates that ripple to both SDKs) land as single atomic commits.

backend/         Go HTTP service. Ingests telemetry, runs the seven
                 failure-class detectors, dedupes into failure_groups,
                 serves the dashboard, fires webhooks on first
                 occurrence of a new failure class. SQLite today,
                 Postgres migration on the roadmap.
sdk-python/      Python SDK shipped to PyPI as `mesedi`. @wrap +
                 @tool decorators, AsyncShipper, Anthropic auto-
                 instrument, hard-halt with budgets + SSE remote
                 channel. Optional `[langchain]` / `[crewai]`
                 extras for framework adapters.
sdk-typescript/  TypeScript / Node SDK shipped to npm as `mesedi`.
                 Feature-parity with Python: wrap()/tool(),
                 AsyncShipper, Anthropic patch, hard-halt + SSE.
                 `mesedi/integrations/vercel_ai` adapter for Vercel
                 AI SDK's generateText.
branding/        Logo and favicon assets.

The customer-facing marketing site, pricing, signup, and live dashboard live in a separate repository (the Next.js front end at https://mesedi.vercel.app).

Local quickstart

Prerequisites:

• Go 1.22+
• Python 3.9+
• Node 18+
• A local SQLite (bundled, no install)

Start the backend (Terminal 1):

cd backend
go run ./cmd/api

Run a sample Python agent (Terminal 2):

cd sdk-python
python3 -m pip install -e .
python3 sandbox/real_agent.py

Or run a sample TypeScript agent (Terminal 2):

cd sdk-typescript
npm install
npm run build
npm run test:sandbox

Open the dashboard:

http://localhost:8080/ui/

You should see executions, events, failure groups, the playbook surface per failure class, the webhook escalation page, and the API-key management page. Default bootstrap API key for local dev: mesedi_sk_dev_local_only.

Status

All seven failure-class detectors are live, plus failure-group deduplication, hard-halt with local budgets plus SSE remote channel, dashboard with collapse-by-class view, operator Halt button, webhook escalation on first-occurrence, framework adapters for LangChain / CrewAI / Vercel AI SDK, and Tier 1 Playbooks v1 covering every detector-signature shape.

The hosted Cloud version runs on Fly.io. Self-host instructions are in backend/README.md.

Documentation

Each subdirectory has its own README:

• backend/README.md: backend runbook, configuration, schema versioning, rate limiting
• sdk-python/README.md: Python SDK API, framework adapters, hard-halt
• sdk-typescript/README.md: TypeScript SDK API, Vercel AI SDK adapter

License

MIT. See LICENSE. Mesedi is operated by Verdifax, LLC d/b/a Mesedi, a Delaware limited liability company.