Commentary
Human review for agentic software teams

Review agent-generated plans and docs as documents—not diffs.

Commentary turns plans, specifications, ADRs, documentation, and live previews into review workspaces where humans comment on meaning and agents receive structured context for the next revision.

Install Commentary for your coding agent

Start with a public GitHub review. Sign in only when you need private access, comments, sharing, or saved review state.

Implementation plan review
platform / agent-rollout
3 unresolved review decisionsReady for agent
Implementation plan review
Rendered planHuman reviewRevision 3
PreviewLatestChanges
Files
implementation-plan.md
acceptance-criteria.md
runbook.md
Rendered Markdown

Enterprise rollout plan

Clarify launch sequence, customer impact, and owner for each readiness milestone before merge.

Name the rollback owner and define the condition that stops the rollout.

Comment threads
Reviewer

Name the rollback owner and define the condition that stops the rollout.

Agent

Revision scoped to Rollout and rollback.

Unresolved context retrievedRevision ready for verification
Product demonstration

Start with the work already in front of you.

Paste a supported public GitHub PR, branch, repository, file, or folder URL. Commentary opens the rendered artifact first and keeps source tools nearby.

Open a real review

Public GitHub content works read-only. Private content and review actions require authentication.

Open GitHub PRs, branches, repos, files, or folders here. Set up private repo access before opening a private repository.

The bottleneck moved

Generation is fast. Judgment is the scarce resource.

Agents can produce polished plans in minutes. That does not make the assumptions correct, the migration safe, or the acceptance criteria complete.

Review should slow down the decision—not hide it inside a chat transcript.
Documents are not diffs

Use the rendered artifact to understand the decision. Use the diff to inspect the change.

Git diffs remain essential source tools. They are not always the best primary reading interface for broad, agent-generated prose.

Diff-first review

  • Broad rewrites fragment the argument across changed lines.
  • Moved sections look new even when the idea did not change.
  • Non-developer reviewers must reconstruct the final document.

Document-first review

  • The complete plan stays readable from problem to rollout.
  • Comments attach to the selected semantic context.
  • Latest and Changes remain available for precise verification.
A human-owned revision loop

AI writes. Humans review. Agents revise.

Every stage has an explicit artifact, durable feedback, and a clear owner.

  1. 01

    Agent prepares the artifact

    A plan, specification, ADR, document, form, or preview becomes explicit reviewable work.

  2. 02

    Commentary opens the review

    The artifact is rendered in the form people need to understand it.

  3. 03

    Humans review meaning

    Reviewers challenge assumptions, scope, risk, acceptance criteria, and user impact.

  4. 04

    Agent revises with context

    Unresolved feedback returns as structured decisions attached to the affected sections.

  5. 05

    Humans verify and approve

    Latest and Changes support verification while final acceptance remains human-owned.

Agent
  1. Agent opens a reviewable plan

  2. Humans resolve blocking decisions

  3. Agent revises and waits for verification

Reviewer
Start where judgment matters most

Three workflows teams can use today.

Begin before implementation, during Git review, or against the working preview.

Review an implementation plan

Challenge scope, architecture, security, testing, rollout, and rollback before the agent writes code.

Draft and Brainstorming Reviews

Review a Markdown pull request

Read the document as readers will experience it, then inspect source changes when they answer a specific question.

GitHub and Azure DevOps

Review the implemented experience

Select the exact UI element and return route, viewport, selector, and screenshot context to the developer or agent.

Opted-in Live Preview
Agent integrations

One review stage across the tools your team already uses.

Connect Codex, Claude Code, GitHub Copilot, MCP-compatible agents, Git providers, the CLI, HTTP APIs, and opted-in preview applications.

CodexClaude CodeGitHub CopilotMCPGitHubAzure DevOpsCLIHTTP API
Explore integrations
Human review without changing source ownership

Keep provider, preview, and permission boundaries explicit.

Commentary adds a human review layer. It does not turn repository access, preview applications, or provider identity into an open-ended agent permission.

Git remains authoritative

GitHub and Azure DevOps remain the source of truth for repository content and provider permissions.

  • Public read-only access
  • Authenticated write actions

Preview access is opt-in

Live Preview Reviews load customer-owned applications only when the app owner integrates the review SDK.

  • Explicit target context
  • Human-controlled capture

Telemetry stays bounded

Marketing telemetry uses allowlisted events and privacy-safe actor identifiers without raw repository or provider identity.

  • Consent-aware
  • No raw URL sink
Review security boundaries
FAQ

What teams ask before adding a human review stage.

Commentary is designed to fit the workflow without hiding who owns source, access, revision, or approval.

Is Commentary only for Markdown pull requests?

No. Markdown review is the Git-native foundation, but the same human-owned workflow also supports pre-Git drafts, implementation plans, forms, knowledge bases, static HTML, and opted-in live previews.

Does Commentary replace GitHub or Azure DevOps?

No. Git providers remain authoritative for source and repository access. Commentary adds a readable review workspace and synchronizes provider comments where the workflow calls for it.

Can I view a public GitHub review without signing in?

Yes. Supported public GitHub reviews are read-only without sign-in. Authentication is required for private access, comments, replies, refreshes using private credentials, sharing, and review submission.

How do coding agents receive feedback?

Agents can use Commentary skills, plugins, CLI, MCP tools, or HTTP APIs to retrieve review identifiers, unresolved comments, revision state, and the context selected by reviewers.

Who owns final approval?

Humans do. Commentary helps agents participate in revision, but it keeps review decisions, permission boundaries, and acceptance explicit and human-owned.

Make the decision reviewable

Review the plan before the agent builds.

Try a complete implementation-plan review, or open the GitHub work your team already needs to evaluate.