claude-code-bridge

Claude Code Bridge

Use this permission-aware local wrapper. Call claude_health({}) first; proceed only when CLI features and authentication are ready.

Start work

claude_task selects workspace, access, model, effort, turns, and session:

{
  workspace: string, // absolute existing real path
  prompt: string, // 1-100,000 characters
  access?: "inspect" | "write", // default: inspect
  model?: string, // health alias or full Claude model ID
  effort?: "low" | "medium" | "high" | "xhigh" | "max",
  max_turns?: number, // 1-100; default: 20
  session?: // default: { mode: "new" }
    | { mode: "new" }
    | { mode: "resume"; session_id: string }
    | { mode: "cloud_create"; description?: string }
    | { mode: "cloud_attach"; target: string },
  execution?: {
    mode?: "auto" | "sync" | "async", // default: auto
    wait_seconds?: number, // 0-45; default: 45
    timeout_seconds?: number // 30-7200; default: 1800
  }
}

• Default to inspect. Use write only for an explicitly authorized change in an absolute, existing, real Git worktree. It selects Claude acceptEdits, auto-approving file edits and Claude-classified common filesystem commands in the validated workspace; other commands and protected paths retain Claude permission checks. This is not OS confinement. Commits, pushes, network, and other external effects require separate target-scoped authorization; never infer a remote, branch, or account.
• Match model/effort to task difficulty; leave them unset for Claude defaults.
• Use explicit session IDs/cloud targets; never infer “the most recent.”
• cloud_create is schema-only and returns unsupported-session-mode; create interactively in Claude Code, then use explicit cloud_attach.
• Inspect uses bounded read/search and plan permissions. Chrome/nested MCP remain disabled; never request bypasses.

After inspect, authorized edits require a new write task; resume with the captured claude_session_id.

Manage jobs

• claude_job_status: { job_id }
• claude_job_result: { job_id, cursor? }; follow next_cursor until absent. Each page is bounded.
• claude_job_continue: { job_id, prompt, execution? }. It preserves the captured workspace, access, model, effort, maximum-turn ceiling, and explicit session. Never add escalation fields.
• claude_job_cancel: { job_id }; cancel only queued or running work.
• claude_job_forget: { job_id }; forget only terminal work. It deletes bridge metadata/output, not Claude's own transcript.