One Terminal. Your Best Models, Working as a System.

Most coding agents make you pick a model and rebuild your context around it. Alloy separates the harness from the model. Use one model directly, route a task by role, or combine two independent proposals, all without changing terminals.

MIT LicensedFree ForeverPre-ReleaseNode 22.19+Built on Pi

View on GitHub →See How It Works

Alloy is the engine we build on. It is free, open, and yours to run.

Why It Exists

No Single Model Is Best at Everything.

One model reasons well about architecture. Another is faster at mechanical work. A third catches what the first two missed. Locking your workflow to one provider means accepting that provider's worst day as your ceiling. Alloy treats the model as a swappable part, because the harness is the thing worth owning.

One Shell

Switch among connected Claude, Codex, and Grok routes without rebuilding your workspace.

Context That Compounds

Project and user facts survive new sessions and new days. Approved skills become reusable operating knowledge.

Boundaries That Hold

Child workflows add credential isolation, budget ceilings, and inherited policy limits.

Three Ways to Work

Chat, Fusion, Auto.

Three modes for three kinds of problem, from a quick answer to a bounded build that reviews its own work.

chat

Fast, Direct Work

You want an answer or an implementation from the active model. Runs the Pi agent loop behind Alloy's shell with memory, skills, MCP, and policy.

/fusion

Two Minds, One Answer

Architect and Builder inspect the same repository independently and read only. A fresh Synthesizer runs after both proposals validate and returns one attributed recommendation. Fusion never writes code.

/auto

Build, Check, Review, Repeat

Checkpoints your repo, prefers an isolated worktree, delegates bounded roles, runs your diagnostics, and asks an independent Reviewer for a verdict before a limited fix loop.

alloy

$ alloy

› /fusion should the journal lock be descriptor bound

architect reading repository, read only

builder reading repository, read only

synthesizer both proposals validated

✓ one attributed recommendation, provenance kept

Safety

Policy Is Code, Not a Polite Request.

Most plan modes ask the model to behave. Alloy enforces the boundary mechanically, so the limit holds whether or not the model cooperates.

Plan and Review: Model tool calls are hard read only. No bash, no edits, no writes, no mutating MCP calls.

Default Approvals: Known dangerous shell patterns, destructive Git, and network or external actions prompt before they run.

Child Policy Ceiling: Child manifests constrain approval profile, sandbox requirement, tools, budget, and concurrency.

Credential Boundary: Child environments are allowlisted. Fusion leases only the selected provider credential into an ephemeral home.

Project Trust: Project config cannot weaken the operator approval profile, sandbox controls, MCP enablement, or budget ceilings.

Host mode is not filesystem or network isolation. Read the security model before running Alloy on untrusted code.

Where Alloy Fits

The Engine Is Open. The Operation Is the Product.

Alloy is the layer our own engineering runs on, and we gave it away. KylairaOS is the layer above it: the memory, the oversight, the approvals, and the day-to-day operation of a real business. You can take Alloy and never speak to us. That is the point. Owning your harness is the only way to stay free of any single model vendor.

Free and MIT

No seat count, no revenue share, no lock-in. Fork it, ship it, keep it.

Provider Agnostic

Your models are configuration, not architecture. When a better one lands, you route to it.

Built in the Open

The same harness runs Kylaira's engineering. What we ship is what we use.

Clone It and Run It.

Alloy is pre-release and moving fast. Issues, forks, and pull requests are all welcome.

github.com/ccoussa717/alloy →See KylairaOS