Agent Residue Evidence MCP server

Agent Residue Evidence

Agent Residue Evidence (ARE) adds an evidence checkpoint to an Agent's test and build workflow:

observe → Agent reviews evidence → user authorizes cleanup
        → Agent cleans with its own tools → ARE verifies

ARE shows task-scoped files, directories, attributed processes, and their listening ports. It never deletes files, stops processes, closes ports, decides that something is safe to delete, scans the full machine, uploads data, uses the network, or collects telemetry.

Visit the project site, run the reproducible task-residue demo, or download the latest verified release.

Standard task

Before the first test or build, the Agent calls begintaskobservation with the exact workspace and any task-owned temporary roots. Events are optional; safe generic events or empty heartbeats can be sent with appendtaskevents. Before its final answer, the Agent calls endtaskobservation, explains the candidates, and asks the user what to do. Only after explicit authorization does the Agent clean with its normal tools. It then calls verifytaskresidue to recheck the same stable candidates.

NOCANDIDATESOBSERVED means only that no candidate appeared inside the declared observation scope. It is not a host-wide cleanliness claim.

Completed task without a baseline

inspectcompletedtask accepts explicit roots and a bounded time window. The result is always PARTIALEVIDENCE; it can use Event, receipt, inferred, or unattributed evidence, but never claims BASELINEOBSERVED. Use getresiduereport to read a saved report without observing again.

Every observation is isolated by an opaque owner capability. Task IDs, report IDs, observation IDs, and candidate IDs are display references only and never authorize access. State is encrypted at rest with opaque filenames; exact paths are revealed only through an authorized candidate-resolution call. Short-lived executor capabilities can append only explicitly allowed event types and root aliases and cannot read, end, verify, or resolve a task.

The MCP surface contains exactly eight tools:

• begintaskobservation
• appendtaskevents
• endtaskobservation
• inspectcompletedtask
• getresiduereport
• verifytaskresidue
• delegatetaskexecutor
• resolveresiduecandidate

There is deliberately no cleanup, delete, execute, terminate, or close tool.

Local-first packaging

One release contains a native CLI/stdio MCP server, thin Agent Plugin, MCPB, checksums, provenance, and SPDX SBOM. See Quickstart and Install. Runtime operation is fully local and offline. Reports default to seven-day retention and a 100 MB total cap; retained reports and active baselines are protected. Uninstall never silently deletes reports.

The current release is published in the official MCP Registry as io.github.fantasyce/agent-residue-evidence.

Native acceptance covers macOS 14+ arm64, Linux amd64, and Windows 11 amd64 through mandatory native CI jobs. Cross-compilation alone is not reported as native acceptance.

Agent Reliability Toolkit

ARE is one independent part of a small, local-first reliability toolkit:

• Agent Recall Trail keeps private Agent memory separate from reviewed shared knowledge.
• Agent Runtime Proof verifies that a live Agent or MCP runtime matches the artifact you approved.
• Agent Residue Evidence records task-scoped files, processes, and listening ports left by tests and builds.

Shortened. The full README is on GitHub.