Browse all docs

Anvil Desktop / Working guide

Operating guide

Set up the workspace

  1. Install and launch Anvil Desktop.
  2. Choose a role so the navigation starts from the surfaces you actually use.
  3. Choose the primary agent for new chats and app-level AI work, then activate any additional providers you want workflows to use. Anvil currently supports Codex CLI, Cursor CLI, OpenAI API, and Azure AI Foundry. On supported Macs, optionally enable Apple Foundation Models for short on-device helper prompts.
  4. Connect delivery providers such as GitHub, Azure DevOps, Linear, Jira, Confluence, Notion, Draw.io, Figma, or Repobase.
  5. Create a workspace.
  6. Add and index local repositories.
  7. Start a conversation from Chat; it is the primary workspace surface.
  8. Start read-only before asking for edits.

The read-only step is not ceremony. Ask Anvil to explain the current code path with file references before asking it to change code. Ambition is cheaper after context exists.

Developer playbook: implement a scoped change

  1. Open the workspace and confirm the target repo is indexed and on the right branch.
  2. Open the work item or paste acceptance criteria into a new Anvil thread.
  3. Ask for the current code path first.
  4. Ask for the smallest safe implementation.
  5. Review the diff in Git.
  6. Run focused checks.
  7. Run Code Review and Security if the change touches auth, data, permissions, payments, production workflows, dependencies, or release gates.
  8. Prepare the PR handover with what changed, what passed, and what could not be verified.

Keep more than one workspace moving

You can switch workspaces while a conversation or terminal is active. Use the sidebar activity centre to return to running work or a thread waiting for approval or input. When Anvil is in the background, desktop notifications can open the originating workspace and thread directly.

Closing a terminal tab stops that terminal process. Switching workspaces or hiding the terminal panel does not; Anvil reattaches to the process and replays its buffered output when you return.

Review playbook

  1. Inspect modified files and branch state.
  2. Run code review with a correctness, regression, security, and test-coverage lens.
  3. Separate blockers from preferences. Preferences should not be escalated as production incidents.
  4. Fix blockers, rerun focused checks, and record residual risk.

Business analyst playbook

  1. Select the work item.
  2. Start a BA session against the relevant indexed repo.
  3. Compare acceptance criteria with current implementation behaviour.
  4. Record questions, risks, dependencies, feasibility notes, and compliance gaps.
  5. Create follow-up work where a finding needs ownership.

Assurance playbook

Use Security for implementation risks, Dependencies for package risk, CI/CD for release gates, and Data and Compliance when the change touches personal data, retention, accessibility, audit, or policy.

Put unverified steps in the handover. Silence is not a test strategy.