Testable AI workflows for the rest of us.
Most AI tooling assumes you have a platform team and a Kubernetes cluster. Heddle assumes you have a laptop — or a few of them — and a problem to solve. Local-first, privacy-first, on-prem-friendly.
→ getheddle.dev for the full overview.
| Repository | What it is | Status |
|---|---|---|
| heddle | The runtime. Python actor-mesh over NATS. Six shipped workers, Workshop web UI, RAG, councils, MCP gateway. Wire-protocol source of truth. | v0.9.2 — active |
| heddle-sdk | Foreign-language SDKs. .NET and Swift, with NATS adapters. Write Heddle processor workers in C# or Swift. | In development |
| warp-design | Vision and ADRs for warp — a macOS-first ad-hoc cluster orchestrator. Pre-implementation; production code lands in warp later. |
Design phase |
| heddle-workspace | Two pillars for the family: (1) shared AI-agent guidance — anchors (philosophy, invariants, contract map), skills, subagents; (2) workspace lifecycle — .heddle-workspace.yaml manifest, bin/workspace CLI for bootstrap and sync across machines. |
Stable |
pip install heddle-ai[workshop]
heddle setup # auto-detects LM Studio + Ollama
heddle workshop # opens localhost:8080Pick a worker in the browser, paste any text, click Run. When you're ready to scale, Heddle adds a NATS message bus for production use.
- Solo / SMB first. No platform team required. No Kubernetes prerequisite. The headline user can install a CLI, edit YAML, and run a workflow in a browser.
- Three model tiers. Local (LM Studio / Ollama), standard (Claude Sonnet), frontier (Claude Opus). Pick cost and privacy per step.
- Privacy by default. Workloads tagged private never leave your machines. Knowledge silos and blind audit patterns prevent the same model from reviewing its own work.
- Typed contracts everywhere. Pydantic messages and per-worker JSON Schemas are the only safety net between actors. No untyped dicts on the wire.
- Stateless workers, deterministic router. Horizontal scaling without coordination; LLM-free dispatch in sub-millisecond.
- Baft — Multi-agent analytical engine for the Iran Transition Project.
- Docman — Document processing pipeline (PDF/DOCX extraction, classification, vector search).
Building something with Heddle? Open a discussion in
heddle — we'd
like to hear about it.
MPL 2.0. Modified source files must remain open; unmodified files can be combined with proprietary code.