Most LLM projects focus on model capabilities; the missing piece for real-world automation is the harness that gives a model "hands, eyes and memory." OpenHarness treats that gap as the product: a small, inspectable Python layer that handles streaming tool-calls, permission checks, context compression and coordinated subagents so an LLM can perform repeatable tasks safely over long sessions.
What Sets It Apart
- Agent-as-infrastructure mindset — OpenHarness isolates the harness responsibilities (tool registry, permission checker, memory, MCP client, task/coordinator subsystems) so you can swap providers or models without reengineering orchestration. This means experiments on provider behaviour can reuse the same safety and tooling layer.
- Rich, pragmatic tooling out of the box — 43+ tools (file, shell, web_fetch, MCP, tasking, git workflows) with pydantic-validated inputs and JSON Schema descriptions. For developers this reduces one-off glue code and makes tool calls observable and type-safe.
- Dry-run & governance features — a dry-run mode previews resolved prompts, auth and tool bindings without executing the model or tools, and multi-level permission modes (default/auto/plan) plus path-level rules let teams balance productivity and safety.
- Personal-agent and channel integration — ohmo packages a personal-agent workspace and gateway for Feishu/Slack/Telegram/Discord; it is opinionated for developer workflows (forking branches, writing code, opening PRs) and runs on existing Claude/Codex-like subscriptions.
Who It's For and Trade-offs
Great fit if you want a lightweight, auditable harness to prototype agent workflows, integrate multiple LLM backends, or experiment with multi-agent coordination without building orchestration from scratch. It’s also useful as a testbed for skills/plugins compatible with Claude-style ecosystems. Look elsewhere if you need a managed SaaS product, an end-user polished chatbot UI, or scale to very large production clusters out-of-the-box. OpenHarness is a developer-focused, repo-first project: expect to manage local auth, provider profiles, and runtime hosting yourself. The project prioritizes clarity and extensibility over turnkey hosting features.
Where It Fits
OpenHarness sits between minimal scripting wrappers (direct API calls) and heavyweight agent platforms. Compared with LangChain-style tool wrappers it provides a stronger opinionated agent loop, permission model and multi-agent primitives; compared with full commercial agent services it gives more control and inspectability at the cost of operational plumbing.
