Alfred turns Claude Code, Codex, or OpenCode into a spec-driven engineering team. Named agents plan the work, write code, test it, review each other, and open pull requests. Alfred keeps working without you at the keyboard. You approve risky actions and decide what merges.
Alfred Desktop is the recommended onboarding path. It detects an existing install, starts or connects to the local server, checks GitHub and engine auth, guides repo selection, and lets you choose a roster theme or custom display names. Slack is still where teams discuss and approve plans.
Recorded from the repository's sample-data fixture in light mode. It shows Ask, Work, Code, Agents, and Settings without using an operator account or live repository.
Each role has one job. The default install includes planning, architecture, implementation, review, test coverage, review fixes, cleanup, summaries, and policy-gated merge.
Reads the parent issue or Slack plan, checks the affected repos, gets approval, and files scoped repo-level work for the implementation roles.
[10:30] architect drafted bundle:billing-v2 [10:31] architect posted plan to Slack [10:34] scope approved [10:35] architect opened backend -> web -> mobile tasks
Picks up a scoped task, makes its own copy of the repo, runs the selected coding CLI with the full project in context, and opens a pull request with tests.
[12:04] senior-dev claimed your-backend#247 [12:04] senior-dev worktree ~/.alfred/wt/12042118/ [12:18] senior-dev PR #1247 opened
Reads a Slack request, GitHub issue, or plain-text plan, asks for missing details, then files small tasks a coding agent can pick up on its own.
[09:02] drake refined settings plan [09:02] drake filed your-backend#248 agent:implement [09:03] drake linked acceptance checks
Reads the changes, runs your test suite in its own copy of the repo, leaves comments on the lines that need work, and marks the pull request ready or needs changes.
[12:18] ras-al reviewed PR#1247 [12:18] ras-al filed 2 inline nits [12:19] ras-al labelled needs-changes
Looks at recently changed files, finds the places that need coverage, and opens focused test PRs that stand on their own.
[15:10] bane found low coverage in billing [15:23] bane PR #1250 opened [15:24] bane labelled tests-only
Finds PRs opened by configured implementation roles with unresolved high-priority review comments, applies the fix in a clean worktree, and pushes the update back to the PR branch.
[16:01] nightwing found P1 on PR#1247 [16:08] nightwing pushed fix [16:09] nightwing asked for re-review
Alfred claims selected work, creates an isolated worktree, invokes the selected coding CLI, routes the result through review, and reports each decision.
Alfred claims the next approved task, assigns the configured role, isolates the run, routes review, records evidence, and reports when a decision is needed.
Alfred runs on a Mac or Linux host you control and uses your authenticated Claude Code, Codex, or OpenCode CLI.
You can plan and approve in Slack, the desktop app, or GitHub. Slack messages cannot bypass the configured approval or merge policy.
Alfred stores candidate lessons under ALFRED_HOME. Structural checks and a separate review decide what can affect later runs. Every memory action remains inspectable and reversible.
Alfred can plan a single feature across several repos at the same time. It works the same way inside a monorepo, whether you use Turborepo, Nx, or Cargo workspaces.
Every line is on GitHub at github.com/luminik-io/alfred. Fork it, read it, run it locally, or audit the code.
Alfred keeps reviewed lessons, file touches, failures, and current work under ALFRED_HOME. Structural checks and a separate model review control promotion. You can inspect, reject, retire, or revert each lesson.
Agents and trusted teammates can propose a lesson. Alfred applies structural checks and a separate model review before the lesson can affect later runs.
Setup blockers, timeouts, and repeated test failures all land in one place. The next time something breaks, your agents start from what happened before instead of digging back through old Slack messages.
Embedded SQLite provides the default recalled-lesson store. FleetBrain keeps the operational ledger and review state. Redis remains an optional scale battery.
The same suite of ten repo-convention tasks plus two controls ran twice
through the claude CLI, changing only whether memory was on. It
measures one thing: whether an agent repeats a mistake the fleet already
learned about.
Method and repro.
Start with a Slack thread, a GitHub issue, or a plain-text plan. Batman is the default-theme name for the public architect role: it handles approved multi-repo fan-out, Lucius handles repo-local implementation as senior-dev, and Drake turns smaller requests into clear tasks.
Batman is the default-theme name for the architect role. It reads the parent issue, checks the affected repos, plans the rollout, gets approval in Slack or the local client, and files scoped child issues. Senior-dev, test-engineer, fixer, reviewers, and the merge gate then finish each PR.
From the moment a run starts to the pull request it opens. The same five stages happen every time an agent fires.
The status view lists configured roles, schedules, recent outcomes, and health. The example below shows the full local roster.
Alfred
alfred status output from a real Mac, captured 2026-06-30 after the OSS full-fleet install. This capture used subscription-backed Claude Code and Codex sessions, so Alfred recorded no separate model charge. Batman, Huntress, Gordon, Bane, Nightwing, and the support jobs are loaded from the first install.
Run doctor for setup health, then dry-run one agent to see the path it would take without touching GitHub, Slack, an engine, or your files.