# AgentInspect

> Debug and regression-test TypeScript AI agents from local evidence

AgentInspect captures framework-faithful execution trees, evaluates them with deterministic TraceFacts and TraceContract rules, creates integrity-verifiable Evidence v2, and lets coding assistants inspect the same local facts over read-only MCP—without a collector, account, or default upload.

- Category: The local evidence debugger and trajectory-test toolkit for TypeScript AI agents
- Outcome: See what your agent did. Prove the fix. Keep the evidence.
- Status: Current release: 6.14.2 · schema 1.0 · Node.js 20+ · MIT · actively maintained
- Site: https://agentinspect.vercel.app
- Docs: https://agentinspect.vercel.app/docs/
- GitHub: https://github.com/rajudandigam/agent-inspect
- npm: https://www.npmjs.com/package/agent-inspect
- Machine facts: https://agentinspect.vercel.app/ai/product.json
- Skill: https://github.com/rajudandigam/agent-inspect/blob/main/.agents/skills/agent-inspect/SKILL.md

## When to use

- Debug a TypeScript AI agent run from local JSONL evidence
- Regression-test trajectories with TraceFacts / TraceContract (no LLM judge)
- Produce integrity-verifiable Evidence v2 for a PR or incident
- Let a coding assistant inspect the same local facts over read-only MCP

## When not to use

- Hosted multi-tenant APM / maintainer dashboard
- LLM-as-judge eval hosting as the primary product
- Compliance certification from the library itself

## Four pillars

1. **Capture faithfully** — Framework-aware execution trees and local JSONL you own.
2. **Test behavior deterministically** — TraceFacts, TraceContract, checks, suites, cohorts, CI gates, and test matchers.
3. **Produce portable evidence** — Offline Evidence v2 with integrity verification and share-policy disclosure.
4. **Debug with coding assistants locally** — Read-only MCP over the same TraceFacts, without a collector or hosted trace database.

## Packages

- `agent-inspect` — core + CLI
- `@agent-inspect/langchain` — LangChain / LangGraph
- `@agent-inspect/ai-sdk` — Vercel AI SDK
- `@agent-inspect/openai-agents` — OpenAI Agents JS
- `@agent-inspect/vitest` / `@agent-inspect/jest` — reporters + experimental matchers
- `@agent-inspect/mcp-server` — read-only MCP (`get_trace_facts`)
- `@agent-inspect/redact` — deterministic redaction

## Key docs

- https://agentinspect.vercel.app/docs/getting-started
- https://agentinspect.vercel.app/docs/trace-facts
- https://agentinspect.vercel.app/docs/contracts
- https://agentinspect.vercel.app/docs/evidence-v2
- https://agentinspect.vercel.app/docs/coding-agent-loop
- https://agentinspect.vercel.app/docs/integrations/langgraph
- https://github.com/rajudandigam/agent-inspect/blob/main/docs/WHY-AGENTINSPECT.md
- https://github.com/rajudandigam/agent-inspect/blob/main/docs/DECISION-GUIDE.md
- https://github.com/rajudandigam/agent-inspect/blob/main/docs/NO-EGRESS-POLICY.md

## Safety

No account · no collector · no default upload · metadata-only by default
Metadata-only by default. No hidden upload. Redaction is best-effort, not certification.