What it does
The tool evaluates an agent execution trace against a published evidence-readiness specification and reports whether the trace records what a reconstruction of four failure classes would need: config drift, session isolation failure, cascade contamination and prompt injection. It checks that evidence is present, not that it is true, and it does not determine whether a failure occurred. The reconstruction engine is not in the repository. The specification is a profile of the OTel GenAI semantic conventions with documented extensions.
Engineers instrumenting agents can use it to see which fields and span kinds a trace is missing, by comparing it with a conformant fixture. Reviewers such as insurers and auditors can use it to judge whether an agent system's telemetry would support retrospective loss review. Coding agents can reach it through an MCP server with three tools: checking a trace, getting a requirement and listing a class's requirements. A GitHub Action can run the same check on pull requests in a repository that produces traces, and annotates each incomplete requirement with its missing field.
The package has no runtime dependencies and requires Python 3.10 or later. The evr command reports universal and per-class readiness, returns exit codes from 0 to 3, and can fold a named class into the exit code. Installing the package provides three console scripts: evr is the checker, while evidence-readiness and evr-mcp start the MCP server over stdio, so evidence-readiness is not the checker. Inline JSON sent to the MCP server is processed in memory.
Limitations and requirements
Requires Python 3.10 or later. Evidence presence does not establish authenticity, and a passing result does not mean the agent is safe. The tool does not determine whether a failure occurred. The reconstruction engine is not included and is available for inspection only under the terms in SOURCE_AVAILABILITY.md. The bare uvx invocation of the package name runs the MCP server, not the checker, so the evr name must be given explicitly. The GitHub Action applies to repositories that produce traces. The MCP configuration needs further setup that the source does not specify.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.BrienBounded/evr
Checks agent traces for evidence-readiness: verifies evidence is present, not that it is true.
v0.1.2As shown on registry.modelcontextprotocol.io when we last read it