Alfred documentation

Everything you need to install Alfred, configure a workspace, design agents and skills, and run a regression suite that catches prompt drift before it ships.

What is Alfred?

Alfred is a local-first agent operations runtime. It does not write your agents for you. It does not pick your model. It does, however, handle everything around the agent that humans and teams keep reinventing: permission policy, eval gating, harness portability, skill loading, trace events, and baseline comparison.

The runtime is split into a harness-agnostic core (packages/core) and a family of adapters that compile the same source-of-truth into each runtime's dialect. opencode, Codex, and Pi have executable adapter previews. Claude, VSCode, Kiro, and external AI surfaces are covered by compatibility contracts or Memory integrations.

New here? Start with Install Alfred → you'll be running the eval suite in under five minutes. Then read Architecture to understand why it is shaped the way it is.

Pick a path

I just want to install it

One curl | sh. No manual git clone. The installer detects your active harness, generates a preview, and waits for human approval before writing harness config.

Read the install guide →

I want to configure my workspace

The guided installer previews project config and smart per-agent model defaults in ~/.alfred/models.json. From there you can set permission policy, register tenants, and wire skills.

See the configuration reference →

I want to understand the architecture

Hexagonal architecture + DDD. The core never imports an adapter. Adapters never redefine domain rules. The boundary is enforced by the build itself.

Read the architecture overview →

I want to set up evals & baselines

Every agent, skill, and policy change must run an eval suite and compare against a baseline. Improvements are accepted on numbers.

Set up your first eval →

Three principles, in order

Local-first. If deterministic local work can solve it, a provider is never called. Avoided calls are traced too.
Deny by default. No agent may broaden its own permissions. Escalation is a human decision, recorded as a trace event.
Evals gate everything. No prompt, agent, skill, or policy change ships without a regression-safe eval pass.

Where to go next

Install Alfred — the one-liner.
Install Alfred — guided TUI with preview-first setup.
Architecture — why it is shaped this way.
Agents & subagents — the six-agent panel.
Evals & baselines — regression-safe prompts.