Magnitude launches open-source offline coding agent for local models

• Magnitude shipped a terminal coding agent that keeps model inference on the local machine: the Magnitude launch post says it runs without API keys, can edit files, use the shell, run scripts, and install with npm i -g @magnitudedev/cli.
• The runtime is bundled into the agent: a runtime reply says it uses a Rust inference engine based on llama.cpp, and the local inference docs say Magnitude manages the model server lifecycle inside the app.
• Model setup is hardware-aware: a model-selection reply says Magnitude can use Hugging Face downloads while recommending local models for the machine, and the models docs describe recommendations by model, quantization, and context size.
• The hardware ceiling is explicit: a Mac hardware reply says 32 GB can fit capable smaller models, while a second hardware reply says larger models on M4 or M5 MacBooks may require aggressive quantization.

The live GitHub README is more specific than the launch tweet: verified model configs, memory estimates before loading, hardware-tuned acceleration, and parallel agents retaining full context windows. The reference docs expose /models, /catalog, and /hardware; the skills docs list .claude, .agents, and .magnitude skill search paths. The troubleshooting page has the local-agent gotcha: session logs and events can contain prompts, project contents, paths, and errors.

Magnitude's current GitHub README describes the project as an open-source agent built on local models, with its own inference engine, no token costs, no API keys, and no rate limits. The official site says local models run fully offline after Magnitude and the chosen model are downloaded.

The runtime lifecycle is spelled out in the local inference docs:

• Downloaded: model files are stored on disk.
• Loading: Magnitude is placing the model and runtime data into memory.
• Running: the model is ready to respond.
• Unloaded: the model remains selected and downloaded, but no longer occupies inference memory.

That bundled runtime is the main product bet. Magnitude removes the separate model-server chore and makes loading, unloading, memory checks, context reuse, and tool-format normalization part of the agent surface.

The get-started docs say first launch detects and profiles local hardware, recommends models, then downloads, prepares, and loads the selected model automatically. One hardware reply says recommendations use measured memory and speed, not only a static spec sheet.

The models docs break the selector into concrete checks:

• Processor, memory, system architecture, and acceleration such as Metal or CUDA.
• Short local inference tests that estimate fit and speed without downloading complete model weights.
• Recommendations for Balanced, Best Quality, Fastest, and Lightweight.
• A specific model, quantization, and context size for each recommendation.
• Catalog fields for memory, intelligence, quantization quality, speed, context, and license.
• Magnitude-managed models in ~/.magnitude/models, plus compatible GGUF models found in the Hugging Face cache.

The Hugging Face reply says users can download any model through HF, but Magnitude tries to remove the up-front model research step.

Out of the box, Magnitude can use the shell, edit files, and run scripts, according to the launch post. The skills docs make the extension surface more concrete:

• vercel-labs/agent-browser, for driving a logged-in Chrome browser.
• anthropics/skills/xlsx, for reading and building Excel spreadsheets.
• anthropics/skills/pptx, for building PowerPoint decks.
• anthropics/skills/docx, for reading and writing Word documents.
• anthropics/skills/pdf, for reading, filling, and creating PDFs.

Magnitude's dependency stack is visible in the launch-thread credits: Hugging Face for the model hub, Unsloth for quants, EffectTS for the agent library, Vercel for skills, and NVIDIA RTX Spark hardware for Linux testing.

The offline boundary is narrower once optional capabilities enter. The get-started docs say web search uses Exa through EXA_API_KEY, and the skills docs warn that skills can run commands, install dependencies, open websites, or interact with external services.

Magnitude's local-model work shows up in guardrails around context, concurrency, and memory. The harness reply says it is roughly the same agent harness with a built-in Rust inference engine, local-model tweaks, and no cloud models.

The local inference docs list the runtime behaviors:

• Preserve the configured context for every request.
• Use extra memory for concurrent work only when the machine can support it.
• Run fewer requests at once rather than silently reducing context on smaller machines.
• Reuse matching context already processed by the running model.
• Stop the model before low memory destabilizes the computer.
• Normalize reasoning formats, tool-call formats, chat templates, and conversation-history conventions across local models.

The circuit-breaker reply adds three local-agent-specific controls: a tool-call circuit breaker to prevent doom loops, measures to prevent overthinking, and more conservative context handling for smaller context windows.

The official FAQ says there is no minimum hardware requirement, but more usable memory means larger models and less memory means simpler work. The replies put names and sizes on that gradient.

• 32 GB: capable smaller models such as Qwen 3.6 35-A3B or Gemma 4 can fit, according to the Mac hardware reply.
• M4 or M5 MacBook: Qwen 3.5 9B or Gemma 4 12B should fit for smaller tasks, while larger models get tough without aggressive quantization, according to the M4 and M5 reply.
• 128 GB: an M5 Max MacBook with 128 GB was described as the best consumer setup, while 64 GB was called a no man's land for additional model access, according to the 128 GB MacBook reply.
• Windows: the Windows reply says Magnitude works through WSL, not native Windows, which matches the get-started docs.

Two future integration points surfaced only in replies. The ACP reply says Magnitude plans to support ACP for integration with other platforms.

A separate chat-completions reply says Magnitude will add a standard chat-completions endpoint for users who want to bring their own inference server. That endpoint would sit alongside the default bundled runtime described by the official site, which says Magnitude does not require Ollama or another model server.