Lohra

A self-improving AI agent runtime — persistent memory, self-authored skills, and a declarative multi-agent workflow harness with Claude-Code-grade rigor. Runs headless: CLI, structured orchestration envelope, and an OpenAI-compatible server. No UI required.

pip install lohra          # Python 3.11–3.13
lohra chat "hello"

Four entry points, none of them a UI

What makes it interesting

• Dynamic workflows as inert data: the agent authors a typed DAG (10 node types — agent, parallel, pipeline, loop_until_dry, verify, judge_panel, gate, completeness_check, checkpoint, nested workflow) that an interpreter runs. No agent-authored code is ever executed; escape is inexpressible, not forbidden.
• Failure is never silent: every failure path produces a fault with its cause; run status is honest (complete | degraded | failed | cancelled | paused).
• Never pay twice: content-addressed per-cell cache — a resumed run replays completed work at zero token cost, across process restarts.
• Human in the loop: checkpoint nodes pause a run until a person answers — in another terminal, another process, another day. Durable state + single-winner leases.
• Cost control: token budgets with soft pre-spawn gates, quota pauses with auto-resume, per-node model/effort/provider routing, operator-owned model tiers.
• Self-improving: persistent memory, self-authored skills, and a workflow library that turns clean runs into reusable templates and bad runs into recorded priors.
• Leaf sandbox: filesystem allowlist (ro/rw), egress allowlist, and taint tracking — operator policy, never the spec.

Configuration

State lives in ~/.lohra (or per-workspace via --profile): .env (API keys) · workflow_policy.json (leaf fs/egress) · workflow_tiers.json (model tiers). Providers out of the box: Anthropic, OpenAI, OpenRouter, DeepSeek, Groq, Together, Gemini, Ollama — plus an opt-in subscription mode (see the ToS warning in lohra auth).

MIT license. Alpha software — built and validated live, but young.