Releases

The first DAGOS release: a local-first DAG operating system for LLM coding workflows.

Core: durable project DAG in SQLite; Jev classifies context (and only classifies); a versioned IR is the only thing providers see; responses are validated before anything becomes durable; every step is an ordered event.
Providers: OpenAI-compatible endpoints (OpenRouter, OpenAI, Ollama, ...) with native tool calling, TypeSafe's Jev through the Decisions API, and an offline fallback so runs never depend on a model Jev.
Context: managed by relevance, not capped. Jev picks nodes, recalls relevant turns from any chat, and leaves out stale tool results; each step is fitted to the model's window; "load all context" works in plain words.
Tools: MCP servers under per-tool policies (off / ask / allow), approvals in the chat, a Jev risk guard that only ever asks more, meta-tool policy inheritance, and Stop.
Semantic review: plain-English lint rules judged per function; a bounded review loop that hands findings back to the model; the Lint view; dagos lint for CI.
App: dagos serve (loopback only) with chat, inspector, DAG panel, Lint view, and settings.

See the README to get started, and specs/001-core-runtime for the specification, design decisions, and traceability.

Navigation Menu