Skip to content
Alfred shield logo

Alfred

Claude Code, Codex, or OpenCode plan work, write code, test it, review changes, and open pull requests. Alfred keeps working without you at the keyboard. You approve risky actions and decide what merges.

Alfred coordinates authenticated Claude Code, Codex, and OpenCode CLIs as short-lived engineering roles. A planner scopes work. A developer implements it in an isolated git worktree. Separate roles review the diff and add tests. The result is a pull request with evidence that a person can inspect.

The host operating system schedules each run. GitHub stores the shared work state. Alfred keeps local run events, lessons, and reliability data under ALFRED_HOME.

Alfred does not replace a coding CLI or proxy model traffic. It adds the workflow, memory, isolation, and decision controls needed to operate several CLI sessions as one autonomous engineering team.

flowchart LR
    request["Request or issue"] --> plan["Plan"]
    plan --> gate["Approval"]
    gate --> build["Isolated build"]
    build --> review["Review and tests"]
    review --> fix["Fix valid findings"]
    fix --> pr["Reviewed PR"]
    pr --> merge["Human or repository policy"]

Short-lived runs

launchd on macOS or systemd --user on Linux starts each firing. No central orchestration loop must stay alive.

Worktree separation

Roles that change or review code use dedicated git worktrees, with locks and recovery around the lifecycle.

Explicit approval

Planned work waits at the configured gate. Alfred does not merge its own PRs by default.

Independent review

The agent that reviews a change is separate from the agent that wrote it.

Local memory

Embedded SQLite recalls reviewed lessons. FleetBrain stores local operations, review, and reliability data.

Inspectable evidence

Run events, checks, review findings, and PR verification remain available for inspection.

  • Claude Code, Codex, and OpenCode engine routing with authentication and readiness checks.
  • Stable roles for planning, architecture, implementation, review, testing, fixes, triage, end-to-end checks, and operations.
  • Single-repository and approved multi-repository planning flows.
  • GitHub issue claiming, branch and worktree management, PR creation, review, and follow-up handling.
  • Embedded SQLite recalled-lesson memory and the FleetBrain operational ledger.
  • Built-in output compaction, structure-first reads, memory budgeting, and blast-radius context.
  • A pinned codebase-memory MCP battery, plus opt-in memory, compression, vector, and graph integrations.
  • CLI, loopback API, browser interface, Tauri desktop app, and optional Slack collaboration.
  • macOS launchd and Linux systemd --user deployment.
  • One operator or small team, one trusted host, one local configuration.
  • Local coding CLI authentication, not a hosted model gateway.
  • Explicit repository scope for scheduling and indexing. This is not a filesystem or network sandbox.
  • Short-lived scheduled processes, not an always-running orchestrator.
  • Human or repository-policy merge authority by default.
  • Optional integrations must be removable and must preserve user-owned config.

Alfred runs with the local user’s permissions. Use a dedicated user, virtual machine, or container when other files or network destinations must be outside the agent boundary.

Read the roadmap for active engine, evidence, memory, and launch work. Read the changelog for shipped changes.

Release tags and package downloads are on GitHub Releases. Claude Code, Codex, and OpenCode are validated execution engines. Any other detected CLI remains unsupported until it passes the same authentication, permission, isolation, and event-contract checks.

License: MIT.