The chamber stays dormant until a local or API model is attached.
Local Relay. Cloud APIs. Anthropic Messages. OpenRouter. One agent loop routes reads, edits, and tool calls through any OpenAI-compatible endpoint.
26 model families parsed natively. Qwen XML, Hermes JSON, Llama Pythonic, DeepSeek, Mistral — Hamr extracts tool calls from raw model output without vLLM normalization.
Budget-aware context compaction. Observable runtime states. Verification lifecycle tracking. Failures are surfaced, not buried in noise.
Hamr is a local-first coding agent built for developers running open-weight models on consumer GPUs. It gives local, open, and low-cost models a stateful runtime — containment, tool execution, verification, and failure handling — plus native tool-call parsers that keep you independent of server-side normalization. Provider-compatible when you need a fallback, local-first by default.
→ Read the docs