Magnitude — précis

Magnitude is an open-source AI coding agent with local models built in. Apache 2.0 licensed. npm install -g @magnitudedev/cli gives you a terminal agent (plus web and Electron desktop clients) that profiles your hardware, recommends and downloads a model from a catalog, and runs it in-process — no Ollama, no model server, no API keys, fully offline after first install. It also supports custom GGUF models from Hugging Face and OpenAI-compatible endpoints.

The pitch ("works out of the box on any hardware") hides a serious engineering core. Under the hood it is a Bun/Turborepo monorepo, Effect-TS native, split across three layers:

• Clients — cli/ (Ink/React terminal), web/, desktop/ (Electron), sharing state through packages/client-common.
• ACN — the local daemon (packages/acn) hosting the agent runtime, sessions, file ops, and display streams. Clients obtain one canonical daemon via JIT "ensurance" coordinated through a single SQLite owner row — no resident coordinator process.
• ICN — the Inference Control Node (inference/), a Rust workspace built on a pinned fork of llama-cpp-rs/llama.cpp, exposing an OpenAI-compatible HTTP API.

The agent runtime itself (packages/agent) is event-sourced: a durable event log is the source of truth, and every screen of state — the context window, task graph, compaction state, conversation, display timeline — is a projection, a replayable query over committed history. Workers own effects; projections own derived truth. This is the single biggest architectural bet, and the thing that separates Magnitude from every other coding agent's imperative in-memory state.

Key techniques worth the read: context compaction as an event-sourced convergence process (bounded 4,096-token summaries, lossy elision as a deterministic liveness fallback), addressed projection state (small indexes in memory, large bodies in addressable entries, residency = exactly the pinned set), display views that page by reshaping rather than slicing, a single-executor inference scheduler with decode-first batching and exact-prefix KV reuse, speculative decoding (MTP/DFlash/DSpark), and hardware-derived memory accounting that separates where an allocation lives from how much it needs.