Objective

Build a single binary, wisp, that is an on-device, tool-using AI microharness on top of Apple’s Foundation Models framework. It uses the same system language model that Apple exposes to end users through the fm command family, but adds what fm lacks: user-defined tools the model can call.

Definition

• On-device: by default every inference runs on the local Apple silicon model with no network calls for generation; Apple’s Private Cloud Compute is an explicit per-session opt-in (ADR 0013).
• Tool-using: the model can request tools by name with typed arguments, the harness executes them, and the results feed back into the model’s reasoning until it produces an answer.
• Microharness: the smallest correct agent loop, not a framework. One session, a registry of tools, a CLI. Everything else is deliberately out of scope until a concrete need appears.

It has two faces:

• CLI: wisp "<prompt>" for a person or script, in the shape of fm respond.
• MCP server: wisp mcp so that other agent harnesses can delegate self-contained work to the on-device model: running a build or test, summarising or classifying text, searching documents, and other non-complex tasks that suit local processing.

Success criteria

1. wisp "<prompt>" answers using the on-device model and calls registered tools when they help.
2. An MCP client can hold multi-turn conversations with the model on named threads over stdio, with wisp’s tools available to the model.
3. Adding a tool is a single Swift type plus one line in the registry; heavier tools are separate binaries (Rust where it suits) that the harness describes to the model.
4. The CLI feels familiar to an fm user: the same flag names where the semantics match.
5. The codebase is exemplary: see engineering.md.

Non-goals (for now)

• Third-party models.
• MCP transports other than stdio.
• Thread persistence across server restarts.
• A plugin system, scripting language, or configuration files for tools.
• A GUI.

Platform

macOS 27 or later, Xcode 27 or later. See ADR 0002.