Open Source
Open work, auditable by anyone.
Free tools, evals, and templates. No key, no account. The same rigor as the paid harness, out in the open so you can read every line.
Self-Inspect
One keyless tool. Your agent sends a thought; it gets back one metathought, a short abstract question that makes it inspect its own assumptions before it continues. Selection is a transparent heuristic over an open CSV. No model, no key.
We measured it
In a 30-turn build (Claude Sonnet 4.6), agents calling Self-Inspect each turn surfaced ~3.5x more assumptions and edge cases than the identical agent without it. It surfaces the forks your agent would otherwise pass over in silence. A thoroughness and transparency amplifier, not a correctness booster, and the full data and a one-command reproduction are public.
More open work
ejentum-mcp ↗
The MCP server for the reasoning harness. Four cognitive modes (reasoning, code, anti-deception, memory) as agent tools. On npm and the official MCP registry.
agent-teams ↗
Multi-agent team templates, each augmented by the harness, plus the eval instruments we publish. Import a workflow, see the diff, run it on your own prompts.
Under Pressure (paper)
Our research on agent reasoning under adversarial pressure. Open preprint with a DOI, free to read and cite.
Everything else ↗
Integrations, skill files, and field notes. All of our open work lives at github.com/ejentum.
The open front door. The full harness is the API.
Self-Inspect is one free question. The Ejentum reasoning harness is the whole cognitive scaffold: four modes, hundreds of engineered abilities, one API call.