Project · 2026

Harness

A self-hostable agent platform that answers with receipts. It checks every claim before it settles.

View on GitHub

Role
Creator
Years
2026
Status
Open source
Links
GitHub ↗

What it is

An open, self-hostable agent runtime and web UI where reliability, verifiable answers, provider freedom, and governance are the core, not add-ons.

What it does

  • Durable runs. Every run checkpoints after every step. Kill the server mid-run, restart it, and the run resumes where it stopped.
  • Verifiable answers. Each claim is checked against the retrieved evidence and labeled grounded, stale, unsupported, or contradicted.
  • Control. Per-run cost budgets, allow/ask/deny tool permissions, a full trace of every step, and an audit log.
  • Any model. Anthropic, OpenAI, Gemini, and any OpenAI-compatible provider, plus MCP tools.

62 backend tests. Built with Python, FastAPI, Next.js, SQLAlchemy, and Postgres.