OpenInterpretability
  • Ekbasis
  • Guide
  • Research
  • Lab
  • Tools
  • Notes
  • Registry
  • Manifesto
Download the modelDownload
OpenInterpretability

Ekbasis: know what an action will do before your agent runs it. From an independent lab for AI-agent safety. Open weights, Apache-2.0: download it and run it yourself.

Download the model
Ekbasis
  • Getting started
  • Download (Hugging Face)
  • For agents (agents.md)
  • Results and models
  • Client and CLI
  • Cookbook
The lab
  • About the lab
  • Open tools
  • Observatory
  • ProbeBench
  • InterpScore
  • Academy
Research
  • Manifesto
  • Roadmap
  • Papers & posts
  • Docs
Community
  • GitHub org
  • Notebooks
  • SDK (openinterp)
  • HuggingFace
© 2026 OpenInterpretability — Apache-2.0 for code, CC-BY 4.0 for docs.Built in public.

openinterp / start

First Result in 10 Minutes

Connect your favorite agent harness (Claude Code, Cursor, Cline) to a Colab session and run probe-causality experiments by conversation. No GPU on your laptop. No data leaves your compute.

  1. 1

    Open the Colab template

    A single notebook with three cells. Free-tier compatible for Qwen2.5-7B at 4-bit; Pro recommended for bf16.

    Open in Colab
  2. 2

    Add your ngrok token to Colab Secrets

    Colab → 🔑 Secrets in the left sidebar → add NGROK_AUTHTOKEN.

    Get a free token at dashboard.ngrok.com

    Secrets stay in your Colab account. We never see them.

  3. 3

    Run the three cells

    The third prints a line that starts with /colab-attach https://…ngrok-free.app. Keep the notebook tab open while you work.

  4. 4

    Install the MCP server on your laptop

    pip install "openinterp-mcp[server]"
  5. 5

    Wire into your agent

    For Claude Code, edit ~/.claude/mcp.json:

    {
      "mcpServers": {
        "openinterp": {
          "command": "openinterp-mcp",
          "args": ["serve"]
        }
      }
    }

    Cursor: ~/.cursor/mcp.json · Cline: settings panel · same JSON.Full guides

  6. 6

    First experiment — paste in your agent

    /colab-attach https://abc123.ngrok-free.app
    # the agent connects, prints loaded model + probes

    Capture L20 at end_question for "Solve x^2 = 4".
    Then run causality_protocol on saturation-direction-L20 with labels [1].

    Total time from Colab attach to first verdict: ~30 seconds.

What you can do next

  • Read the architecture — privacy model, what lives where, why no hosted inference.
  • Browse papers — methodology and findings driving the toolkit.
  • The methodology behind the protocol — paper-6, the three mandatory causality checks.

Your model, your compute, your keys. We just route the conversation.