Docs
Everything you need to ship with Mesedi.
Mesedi observes agent executions in production, detects when an agent is going wrong, 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 the first time a new failure class appears. Start with the Quickstart and the SDK for your stack; everything else is in the sidebar.
5 min
Quickstart
Install the SDK, wrap your agent, see the first execution land in the dashboard. The fastest way to confirm Mesedi is observing your workload.
Python
Python SDK
@wrap and @tool decorators, AsyncShipper, Anthropic auto-instrumentation, hard-halt with budgets, and the framework adapters.
TypeScript
TypeScript SDK
wrap() and tool(), AsyncShipper, Anthropic patching, hard-halt + SSE remote channel, and framework adapters.
HTTP
API reference
Authentication, the executions and events ingest endpoints, schema versioning, rate limiting, and the webhook delivery contract.
Concepts
Failure classes and playbooks
Every detector signature Mesedi ships (twenty-plus and growing), what it looks at, and the canonical playbook each one renders alongside the failure group.
Framework adapters
Wire Mesedi into an existing agent framework without reshaping your code. Each adapter ships behind an optional peer dependency and byte-for-byte matches the wire format hand-instrumented Mesedi produces.
Adapter
LangChain
MesediLangChainCallbackHandler for @langchain/core. Attach to any runnable's callbacks: slot; llm_call and tool_call events flow into the surrounding wrap().
Adapter
LangGraph
instrumentLangGraph(graph) one-call setup. Checkpoint events at every node entry / exit and agent_handoff events on sub-graph invocations.
Adapter
OpenAI Agents SDK
MesediRunHooks implements the RunHooks interface. Checkpoint on agent enter / exit, agent_handoff on transfers, and tool_call per tool.
Adapter
Vercel AI SDK
wrapGenerateText is a drop-in replacement for Vercel's generateText. One llm_call per step, one tool_call per tool invocation, multi-step ReAct native.
Adapter
Mastra
MesediExporter plugs into Mastra's Observability config. No wrap() required — Mastra's root spans own the execution boundary. Emits full run + tool + MCP telemetry.
Notifications
Outbound webhooks fire when a failure group is created (and, optionally, when it recurs). Native adapters cover Slack Block Kit, Discord embeds, and PagerDuty Events API v2.
Reference
Webhooks overview
Event names, payload shape, HMAC verification, per-class filter, severity filter, recurrence modes (off / every_event / throttled), retry policy, delivery log.
Setup
PagerDuty
Four-step walkthrough: PagerDuty service, Events API v2 integration, regional endpoint, integration-key placement in Mesedi. dedup_key behavior + severity mapping explained.
Prefer the source?
Every line of the backend, both SDKs, and the playbook content is on GitHub under MIT license. If you'd rather read the code than the docs, the repo is at github.com/mesedi-ai/mesedi. Each subdirectory has its own README, which is what most of these docs pages are built from.
Questions or corrections? Open an issue on GitHub or email hello@mesedi.ai.