Keep your agent's knowledge linked to your code.
Give important facts an owner, a status and links to the code they describe. Coding agents cite them through MCP, and the GitHub Action flags them when that code changes.
Change SESSION_IDLE_HOURS from 24 to 168 in crates/auth/src/config.rs.
adoc_impacted_by · crates/auth/src/config.rs
Changed paths: 1 path(s)
- crates/auth/src/config.rs
Impacted: 1 object(s)
- auth.session.idle-timeout (claim, verified, owner: team-auth)
via impacts_path: crates/auth/src/config.rs
Proof obligations: 1 obligation(s)
- auth.session.idle-timeout: review impacted claim (evidence: source_code)
adoc_why · auth.session.idle-timeout
Object: auth.session.idle-timeout
Kind: claim
Status: verified
Owner: team-auth
Verified: 2026-08-12
...
Source: docs/agentdoc/sessions.adoc:6:1
AgentDoc links this file to auth.session.idle-timeout, a verified claim owned by team-auth. It says a session ends after 24 hours without a request (docs/agentdoc/sessions.adoc:6).
Stopping before the edit: the claim says 24 hours and this change makes it 168.
Should I make the change and update the claim in the same PR, or check with team-auth first?
Illustrative replay on a sample project. Real adoc output; written agent narration.
MIT licensed. Works with Claude Code and Codex. Alpha downloads for macOS Apple Silicon and Linux.
Watch the 60-second filmKeep the context with the code.
A claim is a documented fact with a status and declared evidence. Add one where a rule needs to be explicit, then make it available to your agent.
Claude Code reads the claims
In this replay, Claude Code reads the claim through AgentDoc and asks before changing the timeout.
How long does a session last before the user is signed out?
adoc_why · auth.session.idle-timeout
Object: auth.session.idle-timeout
Kind: claim
Status: verified
Owner: team-auth
Verified: 2026-08-12
Statement:
A session ends after 24 hours without a request. `SESSION_IDLE_HOURS` in
`crates/auth/src/config.rs` holds the value, and `Session::touch` extends the
deadline on every authenticated request.
...
Source: docs/agentdoc/sessions.adoc:6:1
24 hours without a request. Every authenticated request moves the deadline, so an active user stays signed in.
auth.session.idle-timeout: verified claim, owner team-auth, verified 2026-08-12, docs/agentdoc/sessions.adoc:6
The same claims from the terminal
Find the same knowledge from your terminal, with references to the source file and line.
adoc why auth.session.idle-timeout
Object: auth.session.idle-timeout
Kind: claim
Status: verified
Owner: team-auth
Verified: 2026-08-12
Statement:
A session ends after 24 hours without a request. `SESSION_IDLE_HOURS` in
`crates/auth/src/config.rs` holds the value, and `Session::touch` extends the
deadline on every authenticated request.
Evidence:
- source_code: crates/auth/src/session.rs
- test: crates/auth/tests/session_timeout.rs
Source: docs/agentdoc/sessions.adoc:6:1
Relations:
- depends_on: auth.session.cookie-only
- depends_on: auth.session.server-side
✓ rendered from docs.graph.json
What adoc build writes
Build readable docs and structured data from the same files.
adoc build
docs/agentdoc/operations.adoc:5:1: warning[schema.policy_review_overdue] active policy `ops.production-db-access` is overdue for review (last effective 2026-06-01, review interval 90d, due 2026-08-30)
object_id: ops.production-db-access
help: Re-review the policy and update `effective_at`, or adjust `review_interval`.
docs/agentdoc/uploads.adoc:35:1: warning[lifecycle.expired] Knowledge Object `uploads.malware-scan` expired on 2026-09-01.
object_id: uploads.malware-scan
help: Update `expires_at` or remove it if this Knowledge Object is still valid.
info[build.embeddings_cached] embeddings: cached 15, computed 0
0 errors, 2 warnings
ls docs/agentdoc/dist
docs.graph.json
docs.html
docs.search.json
Illustrative replays on a sample project. The adoc output is real; the agent narration is written for this page. Expiry counts were captured on 2026-09-26.
People mark claims as verified. AgentDoc checks their structure and declared evidence. Read the source format (opens in a new tab).
- Changed code
- crates/auth/src/config.rs
- Linked claim
- auth.session.idle-timeout
- Owner
- team-auth
Illustrative report excerpt using the sample claim shown above.
Review knowledge when the code changes.
The GitHub Action flags verified claims linked to changed code and reports changed paths without knowledge coverage.
Reviewers can see what needs a look alongside the code change. The report needs no Claude credential and is advisory by default.
Optional Claude drafting proposes new claims for uncovered code and sends selected changes and claims to Anthropic. Drafts still need human review.
Governed knowledge for humans and agents.
Cloud keeps one reviewed record of what your team knows. It reads GitHub, Slack and Confluence, sends each change to the person who has to decide, and shows what every agent read and when access was denied. The CLI, MCP server and GitHub Action work without it.
One repo, with your agent and your pull requests.
AgentDoc Cloud. Governed knowledge for humans and agents.
One reviewed record of what your company knows.
AgentDoc in 60 seconds.
A silent walkthrough of a claim, Claude Code reading it over MCP and the GitHub Action on a pull request, then what comes next.
No sound. Illustrative replay on a sample project. Real adoc output; written agent narration. The pull request is illustrative.
Read the film transcript
- 0:00Your coding agent does what the docs say. docs/sessions.md: A session ends after 24 hours without a request. crates/auth/src/config.rs changes SESSION_IDLE_HOURS from 24 to 168. Nothing tells it the docs are out of date.
- 0:07AgentDoc Keep your agent's knowledge linked to your code.
- 0:11Claims. Give a fact an owner, a status and a link to the code it describes. Claim auth.session.idle-timeout: status verified, owner team-auth, linked code crates/auth/src/config.rs. adoc check: 0 errors, 0 warnings. People mark claims verified. AgentDoc checks their structure and declared evidence.
- 0:20Claude Code over MCP. With the claim in front of it, Claude asks before the edit. Prompt: Change SESSION_IDLE_HOURS from 24 to 168 in crates/auth/src/config.rs. Claude calls adoc_impacted_by and adoc_why, which return the claim auth.session.idle-timeout. Claude: Stopping before the edit: the claim says 24 hours and this change makes it 168. Should I make the change and update the claim in the same PR, or check with team-auth first? Illustrative replay. Real adoc output, written narration.
- 0:32GitHub Action. When linked code changes, the pull request says so. AgentDoc PR Report: Knowledge review needed. 1 proof obligation needs a decision. team-auth: decide on auth.session.idle-timeout. Its source changed in this PR while verified. Either re-verify or set status: draft until reviewed. Claude can draft the update. A person decides what goes in. Optional Claude drafting, experimental: 1 knowledge update proposed. Review it below and apply what is right. Nothing was committed. docs/agentdoc/sessions.adoc: A session ends after 168 hours without a request. Would move verified to draft when applied. A team-auth reviewer approved these changes. Illustrative pull request.
- 0:42Today, open-source alpha: one repo, with your agent and your pull requests. Next, in development: AgentDoc Cloud. Governed knowledge for humans and agents. Later, vision: one reviewed record of what your company knows.
- 0:51The more your agents know, the more you can let them do. More autonomy. Without losing control.
- 0:57AgentDoc. Open-source alpha. MIT licensed. Works with Claude Code and Codex. agentdoc.dev
Try it on one repo.
Start with an existing doc and follow a citation back to its source. Add a typed claim when you need a review status or a link to code, then connect your coding agent.
Free and MIT licensed. No Cloud account needed. The guide covers the alpha's supported platforms and setup steps.
Questions about AgentDoc
The answers describe the alpha as it works today.
Do I need to rewrite my docs?
Plain Markdown is searchable and citable. Typed claims add status, ownership and declared code links in AgentDoc Source files. Start with one. Keep AGENTS.md and CLAUDE.md for instructions.
Does AgentDoc verify that a fact is true?
No. A person reviews and marks a claim verified. AgentDoc checks its structure and declared evidence. The GitHub Action flags verified claims linked to changed code.
Does my knowledge reach a model?
The CLI and MCP server make no model calls themselves. MCP returns knowledge to your coding agent, which may send it to its configured model provider. Optional Action drafting sends selected changes and claims to Anthropic. Embedding search downloads a model on first use; lexical search works offline.
Which coding agents work with it?
Claude Code and Codex are confirmed for MCP access. Other clients that can start a local MCP server may work too. Drafting in the GitHub Action uses Claude Code.