Claude

Triage issues, pick up backlog work and review linked pull requests with a local or cloud agent.

Set up an agent

Connect Claude from Settings → Integrations → Claude. Enable a Claude Code worker, a local API agent, or the cloud agent. Automatic workflows are off by default. Workspace owners and administrators can configure them under Autonomous Claude.

The Claude Code worker edits, tests, and commits code in the writable repository you select. It uses your installed Claude Code sign-in. Local API and cloud agents handle workspace tools and GitHub actions; those API runtimes need a separate executor to edit repository code. Opening a pull request requires pushed commits. Configure repository mappings and templates in the GitHub integration.

Start a Claude Code worker

Install Claude Code and sign in with claude auth login. Use Node 18.20 or newer. Install the Velocity worker, sign in to Velocity, and select one or more workspaces. The worker saves private credentials and refreshes them automatically:

npm install --global https://github.com/velocity-quest/velocity-agent/releases/download/agent-v0.5.2/velocity-agent-0.5.2.tgz
velocity-agent login
velocity-agent whoami
velocity-agent run --executor claude-code --repo /path/to/repository --team YOUR_TEAM

For SSH or a browser on another machine, use velocity-agent login --no-browserand paste the login code shown by Velocity. Input is hidden. You can also paste the full redirected URL when a local callback cannot arrive. Use velocity-agent logout to revoke the saved grant. For token-based automation, see the MCP setup guide. No workspace flag is required after browser login.

The worker polls all approved workspaces by default. Add --workspace ID_OR_SLUG to focus on one. Sign in again to expand an older single-workspace grant. The worker claims queued Claude Code runs for the selected team. Add --allow-delivery to authorize pushing and opening a PR. Connecting Claude Code or Claude.ai to MCP tools alone does not start a worker. Both clients use the same endpoint; see client setup.

Automatic workflows

  • Triage new issues: processes issues created while this mode is enabled. It preserves existing labels, adds suitable labels and suggests priority in a comment. It can set priority when the issue currently has None. Optional assignment assigns Claude only when neither a human nor another agent is assigned.
  • Pick up backlog issues: selects unassigned issues by priority, then age. It assigns Claude and moves the issue into progress when that team has an In Progress status. It reports tasks that need a code executor instead of claiming code was run.
  • Review linked pull requests: reviews newly linked, open PRs in the issue’s configured GitHub repository. It reads the current commit and changed-file patches and posts a GitHub COMMENT review tied to that commit. It does not approve or merge PRs. Reviews cover at most 300 files and a bounded context; missing and truncated patches are identified.

The scheduler runs periodically and persists its scan position. A linked PR gets one initial review even if both a webhook and an MCP call link it. Enabling triage or review starts from that time; it does not process all historical issues or PRs.

Scope and concurrency

Filter by project, label and priority. Empty filters include everything. Every selected dimension must match; any one selected label is sufficient. None is a distinct priority. The agent rechecks scope before acting, so changing filters can cancel runs that are no longer eligible.

The concurrency limit is shared across local and cloud agents in a workspace, from 1 to 10, with a default of 1. Waiting for approval frees an active execution slot. Automatic backlog pickup also reserves pending and waiting work so it does not flood the queue.

Optional action approval

Approval is off by default. Enable Require approval for sensitive actions to pause before creating a branch, opening a PR or moving an issue to a completed or cancelled status. The pending action shows resolved arguments and repository settings. An owner or administrator can approve or reject it from the Claude settings page.

Approval applies to that exact action and policy, expires after 24 hours and is consumed once. The runner saves its remaining tools and conversation, then resumes after approval without repeating earlier tools. Changing policy or repository settings invalidates the action. Rejecting an action cancels its run. Turning approval off resumes waiting runs under the current policy.

This checkpoint is supported by the API workers. The native Claude Code worker requires it to be off; it uses Claude Code's own tool controls and the explicit --allow-delivery flag.

Emergency stop

Emergency stop cancels queued and active managed runs and blocks new automatic and reactive dispatch. Both runtimes check ownership and policy before tools and check for cancellation during model requests every five seconds. Writes already sent to a provider cannot be undone.

Resume Claude allows new eligible work to be queued. Cancelled executions do not continue their old conversations. Velocity stops the Claude Code subprocess owned by a managed worker when its next heartbeat detects cancellation. Independent Claude Code chats in your IDE use their own stop controls.

Run history

Run history records the trigger, runtime, outcome, token usage and cost. A run waiting for approval keeps its checkpoint and usage totals. Pending actions are shown separately with their expiry and preview. If an execution expires or an approved write has an uncertain outcome, inspect the issue and GitHub before starting a new run.