Skip to content
SolenopsisCampo edited this page Nov 8, 2026 · 2 revisions

Venastine Research Harness

An open-source, security-first agentic harness for large language models: a tool-using chat assistant and a ten-pass deep-research pipeline that fact-checks, critiques and scores its own claims, in your terminal or in the browser.

It has two modes that share the same machinery:

  • Chat: a normal tool-using assistant loop.
  • Research: a ten-pass pipeline that writes an answer, then independently fact-checks and critiques its own claims, audits the assumptions hiding underneath them, scores every claim by a fixed formula rather than by asking the model how sure it is, and revises only what came back flagged.

It works with Anthropic, any OpenAI-compatible provider (OpenAI, DeepSeek, Groq, Mistral, local endpoints), and Google Gemini.

This wiki holds guides for people using the harness. The README is the full reference, and AGENTS.md is where to start if you want to work on the code.

Start here

  • Installation: install from GitHub, npm, yarn, bun, pnpm or a clone, then update or remove it.
  • Getting Started: the first launch, connecting a model, and your first chat and research run.

Guides

  • Tool Permissions: every tool the model can call, what it does, whether it is on, and when it asks you first.
  • Shell and Sandbox: turning on the shell, its approval modes, and where each command runs.
  • Writing Agents and Skills: the file format, every field, and worked examples.
  • MCP Servers: connecting third-party tools through the Model Context Protocol.
  • FAQ: common problems and how to fix them.

Try it without installing

The same harness runs in your browser on a Hugging Face Space and a ModelScope Studio. Each tab gets its own session, deleted when the tab closes. Connect a model there with /connect huggingface or /connect openrouter. The shared demo takes no pasted key and does not offer the file and shell tools.

Getting help

  • A bug, a missing feature, or a doc that is wrong: open an issue and pick the form that fits.
  • A security vulnerability: do not open a public issue. Follow SECURITY.md to report it privately.
  • Contributing: read CONTRIBUTING.md first.

These pages were checked against release 0.3.12 (commit f33ef8a). Where a page and the README disagree, the README is the reference.

Clone this wiki locally