Retrace bronze

record-agent-run

Record an AI agent execution as a trace: every model call, tool call, retrieval, and error is captured as a span with its input, output, cost, and timing.

• record this agent execution so I can replay it later
• instrument my Python agent so every model call is captured

inspect-trace

Read a recorded agent execution and walk its span tree to find the first step that failed, drifted, or produced an ungrounded answer.

• show me the span where this agent run failed
• walk the execution tree for trace 6f2c… and find the first error

search-spans

Semantic search across recorded spans and agent memories to find prior occurrences of a behavior, prompt, or failure by meaning rather than exact text.

• have I seen this hallucination before in any recorded run
• find previous runs where the agent looped on the same tool

fork-and-replay

Branch a recorded execution at one span with a modified input, then cascade-replay every downstream step so the whole agent re-executes along the new path. Side-effecting: requires explicit user approval before invocation.

• fork this run at step 3 with a shorter prompt and re-execute the rest
• re-run the agent from the retrieval step using a different model

prove-the-fix

Re-run a failed execution after a change and return a verdict — improved, regressed, or unchanged — with the first divergent span. Side-effecting: requires explicit user approval before invocation.

• did my prompt change actually fix this failure
• re-run the failed trace and tell me if it improved or regressed

evaluate-and-gate

Score recorded executions against behavioral criteria and return a pass/fail gate decision for a CI/CD pipeline.

• score these runs against my criteria and fail the build if they regress
• run the CI evaluation gate for this deployment

enforce-budget

Ask before the next model or tool call whether it is allowed: returns allow, block, or hold based on cost budgets, loop detection, and latency ceilings.

• is this next tool call within the cost budget
• should the agent be allowed to make this call

agent-memory

Store and semantically recall persistent facts, preferences, corrections, and patterns extracted from prior agent runs.

• remember that this customer prefers metric units
• what have I learned about this project from earlier runs

publish-tape

Publish a recorded execution as an interactive, shareable public URL so a human can replay it in a browser. Side-effecting: requires explicit user approval before invocation.

• publish this run as a shareable link for my teammate
• make an interactive replay of this execution public

read-retrace-docs

Search and read the Retrace documentation, SDK guides, CLI reference, and OpenAPI specification over MCP, without credentials.

• how do I enable cascade fork replay
• what are the Retrace rate limits

Retrace MCP Server

Product MCP server for recorded AI agent executions: list and read traces, semantic span search, agent memory, fork replay, evaluation updates, and tape publishing.

Retrace Docs MCP Server

Documentation MCP server: search and read Retrace docs, SDK guides, CLI reference, and the OpenAPI specification. Read-only, no credentials required.

Retrace Agent Card (A2A)

Agent-to-Agent card describing Retrace's skills for recording, replaying, forking, and evaluating AI agent executions.

Retrace REST API (OpenAPI 3.1)

OpenAPI 3.1 description of the Retrace REST API: trace ingestion, semantic search, fork replay, batch operations, evaluation gates, runtime enforcement, and OTLP ingestion.

Retrace Agent Skills

Index of Retrace agent skills, each with a name, description, and entry point.