record, replay & guard every agent turn

Q: How does AgentFlight replay multi-step AI agents in CI with $0 token spend?

FlightRecorder captures every LLM turn, streaming delta, and MCP tool call into a canonical-JSON SHA-256 hash chain (.flight.json or .flight.jsonl) with secrets automatically scrubbed. CassetteReplayer and the built-in pytest plugin stub matching tool calls locally in 0.08 to 2.8 microseconds and raise TrajectoryDivergenceError when arguments drift.

Q: What does the core Python SDK require at runtime?

The core AgentFlight recording, replay, circuit-breaker, diffing, OpenTelemetry export, redaction, and MCP linter engine under src/agentflight runs on pure Python 3.10+ standard library with zero external runtime dependencies.

Q: Can I connect AgentFlight to Claude Code, Cursor, or CI pipelines?

Yes. Run agentflight serve-mcp locally via uvx, connect directly to the public Streamable HTTP endpoint at https://agentflight-tau.vercel.app/mcp, or gate pull requests using agentflight ci and pytest --flight-mode=replay.

the open-source black-box flight recorder, loop breaker & MCP linter

pip install agentflight-mcp

open studio →

$ agentflight demo

FlightRecorder: 4 spans chained (SHA-256 verified, API keys scrubbed)
CassetteReplayer: replayed 4/4 spans in 0.011 ms  [$0.0000 CI spend]
TrajectoryCircuitBreaker: tripped ping_pong_oscillation (A -> B -> A -> B)
ToolSmellLinter: 200 MCP schemas audited in 119.3 ms (1,676 tools/sec)

0.08 us/span replay · 10 MCP rules · pure Python 3.10+ stdlib core

Try MCP & Breaker Studio Read docs

Tamper-Evident

Every LLM turn and MCP call is hashed into a canonical-JSON SHA-256 chain with API keys scrubbed automatically.

$0.00 CI Replay

Replace flaky 30-second live API CI runs with deterministic cassette replay executing in microseconds offline.

Loop & Cost Shield

Stop infinite repeat calls, A↔B ping-pong cycles, error storms, and dollar budget spikes before spending another token.

10 MCP Rules

Lint schemas statically across TS001–TS010 to catch missing descriptions, deep nesting, overlap collisions, and unguarded deletes.

Works Everywhere

Drop into OpenAI, Anthropic, Pytest, Claude Code, Cursor, or OpenTelemetry pipelines — pairs cleanly with Prometheus, StackPulse, Plausible, and Vercel.

SDK Wrappers Pytest CI MCP Server OTel Traces

Inspect verified Python 3.12 benchmarks, test schemas in the interactive studio, or read all 10 TS & 5 breaker rules.

questions

How does AgentFlight replay multi-step AI agents in CI with $0 token spend?

FlightRecorder captures every LLM turn, streaming delta, and MCP tool call into a canonical-JSON SHA-256 hash chain (.flight.json or .flight.jsonl) with secrets scrubbed. CassetteReplayer and the built-in pytest plugin stub matching tool calls locally in microseconds and raise TrajectoryDivergenceError when arguments drift.

What does the core Python SDK require at runtime?

Everything inside src/agentflight (recording, replay, loop circuit breaker, trajectory diffing, OpenTelemetry export, PII redaction, and the stdio MCP server) uses pure Python 3.10+ standard library only.

Can I connect AgentFlight to Claude Code, Cursor, or CI pipelines?

Yes. Run agentflight serve-mcp via uvx for local stdio workflows, connect to the hosted Streamable HTTP endpoint at https://agentflight-tau.vercel.app/mcp, or gate pull requests using agentflight ci.