Skip to main content

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

Port Command For
Human CLI lohra chat you, in a terminal
Orchestration envelope lohra chat --json other agents/scripts — one parseable JSON per turn (input/output/reasoning/tool_calls/usage)
OpenAI-compatible API lohra serve any OpenAI client becomes a Lohra client
WS/REST gateway lohra dashboard optional, only if a UI attaches

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.
  • Live supervision: durable notices that survive turns and dead turns, mid-flight leaf steering, lohra workflow list|watch|audit as a zero-token spectator view, and an explicit agent×human boundary doctrine for autonomous fixes.

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, xAI, Zhipu GLM, Moonshot Kimi, Ollama — plus an opt-in subscription mode (see the ToS warning in lohra auth).

Driving Lohra from another agent? lohra skill export use-lohra --to .claude/skills (or --to .codex/skills) drops the delegation kit — a skill teaching Codex CLI / Claude Code how to hand Lohra self-contained work through lohra chat --json.

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

Metadata

Release files for lohra 0.0.16

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for lohra 0.0.16
File Size Uploaded Provenance Actions
lohra-0.0.16.tar.gz 781.1 kB Without provenance Details for lohra-0.0.16.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for lohra 0.0.16
File Interpreter ABI Platform Actions
lohra-0.0.16-py3-none-any.whl Python 3 none any Details for lohra-0.0.16-py3-none-any.whl

Total release size: 1.3 MB

Release files / lohra-0.0.16.tar.gz

Download URL lohra-0.0.16.tar.gz
Size 781.1 kB
Tags Source
SHA-256 checksum
How to use checksums
4cde7b5a3f95ef3865e77e16728f9386d5b14d8f1263b0b638e6c8ca4ae67b1b
BLAKE2b-256 checksum
How to use checksums
acde8546dbd4d5f583108d9295c030b6828c625c736f225b8deccea375399c34
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.10.17

Release files / lohra-0.0.16-py3-none-any.whl

Download URL lohra-0.0.16-py3-none-any.whl
Size 473.8 kB
Tags Python 3
SHA-256 checksum
How to use checksums
8f39583225c3d3d76f16306df304fb0501102604495213f0636c8e0a8df30e0c
BLAKE2b-256 checksum
How to use checksums
e97b907f0e2a034796e34ffc1749de2784784f4684bd6e71234a2ebc6ba67ef3
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.10.17

Security for lohra 0.0.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page