Skip to content

Model-neutral by design

harnessforge doesn't care which LLM is the brain.

At provisioning time

harness init builds a profile. By default it uses whatever LLM provider you have credentials for:

  • Anthropic (ANTHROPIC_API_KEY + pip install 'harnessforge[anthropic]')
  • OpenAI (OPENAI_API_KEY + pip install 'harnessforge[openai]')
  • Google Gemini (GOOGLE_API_KEY + pip install 'harnessforge[gemini]')
  • Ollama (local; pip install 'harnessforge[ollama]')
  • LiteLLM (100+ providers via pip install 'harnessforge[litellm]')

If none is available, harness init falls back to the deterministic template profiler — no API key required. The output is still production-grade; the LLM only refines descriptions and conventions.

--no-llm forces the template path explicitly. Recommended for CI.

At runtime

We don't run your agent. Your coding agent — Claude Code, Cursor, Codex, Gemini CLI, Aider, OpenHarness — runs whichever model it's configured for. We just generate the files it reads.

At verification time

harness verify runs blueprint validators. Validators are plain Python (no LLM). They check structural invariants (schema, citations, lineage, tool log, idempotency). No model required.

File-format neutrality

The generated files follow the de facto conventions used by every major agent project in May 2026:

  • AGENTS.md — OpenAI Codex CLI standard
  • SOUL.md — Hermes + OpenClaw convention
  • TOOLS.md — OpenClaw convention
  • MEMORY.md — OpenHarness convention
  • SKILLS/<name>/SKILL.md — anthropics/skills format
  • Plus per-IDE files (.claude/CLAUDE.md, .cursor/rules, .continue/config.json, .windsurf/rules)

A single bootstrapped repo is usable by any coding agent that follows any of these conventions. No tool gets second-class treatment.