Agentic Reports rebuilds around evidence you can check

What Agentic Reports is

Most AI reporting tools hand you fluent prose and leave you to trust it. Agentic Reports does the opposite. It selects passages from documents you supply and keeps the source, the capture time and a content hash attached to each one, so a reader can check the claim rather than believe it.

The output is an extractive evidence report, not model authored narrative. You can also compare two reports and see exactly which evidence changed between them.

What changed

Commit 1459a17 revived a project that had been dormant since 2024 and rebuilt it around governed retrieval. Documents stay in the local process unless you explicitly run the optional discovery command.

The commit brings six official SDK MCP tools, a policy resource, a local CLI and subprocess tests with fixture benchmarks. Publication dates are treated carefully: capture freshness does not establish publication freshness, and a publishedAt value cannot be later than capture time.

Get started

Node 24 and npm. Run from a checkout:

git clone https://github.com/ruvnet/agentic-reports.git
cd agentic-reports
npm ci --ignore-scripts
npm test
node src/cli.mjs generate < fixtures/evidence.json > report.json
node src/cli.mjs verify < report.json

Expected result: report.json is written from the synthetic fixture, and the verify command validates it and exits without error.

The MCP server runs from the same checkout, so the config needs an absolute path:

{"mcpServers":{"agentic-reports":{"command":"node","args":["/absolute/path/agentic-reports/src/cli.mjs","mcp"]}}}

The MCP tests only run when the operator sets REPORTS_ALLOW_VALIDATION=1 before starting the server. The discover command is deliberately absent from MCP.

Use it today

Practical case: a policy or security team that needs a defensible summary of a document set. Input is the evidence JSON with source, URL, capture time and text. Workflow is generate, then verify, then html for something readable. Output is a report where each claim points at the passage it came from.

Acceptance test: generate a report, edit one character inside a quoted passage in the JSON, then run verify again. Verification should fail.

Push it further

Experimental commentary. The interesting use is longitudinal: run the same topic weekly and diff the reports to watch a body of evidence move. The compare command accepts a before and after report on stdin.

Limitation: this is a local single operator boundary, not a hosted multiuser service, and receipts carry an unsigned output hash. Falsifiable test: run compare on two reports built from the same fixture. It should report no changed evidence.