Claude Code

developerz.ai is a thin orchestrator, it does not write code. It triages the issue, talks to the reporter, then hands off to Claude Code. The primary path, handoff.mode: fleet, dispatches straight to the developerz-hosted BYOVM fleet; running Claude Code on your own box is still supported over webhook. Claude Code writes the fix and opens the PR; the maintainer agent watches CI and review and merges when the machine gates pass. Coding stays in Claude Code; developerz.ai only runs the loop around it.

How the handoff works

Set handoff.mode: fleet in maintainer.yml:

# yaml-language-server: $schema=https://developerz.ai/schemas/maintainer.v1.json
version: 1

handoff:
  mode: fleet             # webhook | label | fleet | none
  ask_reporter_first: true

mode takes one of four values, default none:

When a new issue is triaged and qualified as a bug or feature, a fleet dispatch writes a durable tasks row that the fleet assigner claims. With ask_reporter_first: true the bot asks the reporter for consent first, and a dispatch with no agreement from the reporter is refused. See the maintainer.yml reference for every handoff field.

Two gates can refuse a fleet dispatch before any work is queued. An account with no AI devs is refused no-capacity, and it clears only by a grant or a seat; a full backlog is queue-full until it drains. The API doc has the exact REST responses.

Run Claude Code on your own box instead? Set mode: webhook and point it at a dashboard webhook wired to your Claude Code runner. webhook_ref is required there and the webhook must be registered in the dashboard; a qualified issue then fires a signed issue.handoff event your runner picks up. See the webhooks doc.

Convention support

Claude Code works from the repo's own steering files, and the handoff lanes read them back rather than only writing prose to a database:

• CLAUDE.md / AGENTS.md brain. Whichever of the two root files exists is read back into the run's prompt; AGENTS.md is treated as the same artifact under the vendor-neutral name, and a run takes only the FIRST hit, never both concatenated. The file is capped at a fixed character budget so a large brain file cannot silently eat the model's working room; past the cap the run keeps the first portion and appends a visible marker naming the true size and the path.
• .claude/agents roster. Specialist definitions placed under .claude/agents/ are indexed by name, path, and one-line description, and routed by matching a unit of work's file paths and purpose against that index. A repo with no roster behaves exactly like one with no .claude/agents directory at all.
• Setup scaffolding. Setup mode can write a starting CLAUDE.md, .claude/agents/, and .claude/skills/ for a repo that has none. It is establish-or-verify: it never overwrites files the repo already has.

What the PR loop does

Once Claude Code's PR is open, the maintainer agent takes over:

1. Waits for CI to go green and review threads to settle.
2. Classifies risk: breaking changes, large diffs, and first-time contributors pause for approval (escalate.ask_before_acting).
3. Merges when the machine gates pass, CI green, review verdict, and branch protections, never blindly. Set auto_merge: false to keep a human in the loop.

BYOK: your key, your bill

Claude Code runs on your Anthropic key. developerz.ai never proxies or resells tokens, your provider invoice shows exactly what the agent spent. Swap models any time without touching your policy. See the BYOK doc.

Native tools over MCP

Claude Code can connect straight to our MCP server at mcp.developerz.ai and use the maintainer tools natively, dispatch tasks, check fleet status, re-run a review, from inside the editor. Add the server to .claude/settings.json with a PAT; the Quickstart has the exact config. MCP is the integration surface; the full tool list is in the MCP verb reference.

What's audited

Every maintainer decision, the handoff dispatch itself, and every MCP tool call Claude Code makes against mcp.developerz.ai, is written to the append-only audit log. The bot always discloses that it is an automated agent, and defers to any other review bot on the PR (no bot-on-bot loops). If it isn't logged, it didn't happen. See the security doc.