Learn

Documentation

Everything you need to build with Ejentum. From quickstart guides to advanced patterns.

Changelog

API changes and updates. For the full development history, see our GitHub.

v1.1. May 2026

Adaptive modes, renamed tools, production gateway.

  • Eight modes. Each of the four harnesses now ships in a dynamic and an adaptive variant: reasoning / adaptive-reasoning, code / adaptive-code, anti-deception / adaptive-anti-deception, memory / adaptive-memory. Dynamic returns the best-fit operation as-is; adaptive rewrites the procedure and topology to your task while keeping the safety guards identical. Adaptive requires the Go or Super tier.
  • MCP tool names are the mode strings. The eight MCP tools are named exactly reasoning, code, anti-deception, memory, and the four adaptive- variants.
  • Production gateway. The REST endpoint is https://api.ejentum.com/harness/; the hosted MCP endpoint is https://api.ejentum.com/mcp (Streamable HTTP, Bearer auth via EJENTUM_API_KEY). The stdio ejentum-mcp npm package works unchanged for Claude Desktop, Cursor, Windsurf, Claude Code, Cline, Continue, and other subprocess-spawning clients.
  • Pricing. Free trial (1,000 dynamic calls, no card), Go (€5/month: 1,000 dynamic + 250 adaptive), Super (€25/month: 5,000 dynamic + 1,500 adaptive).
  • See the updated MCP integration guide and API Reference.

v1.0.x. May 2026

Official MCP server.

  • ejentum-mcp: an MCP server exposing the harnesses as tools any agentic client can call (Claude Desktop, Cursor, Windsurf, Claude Code, n8n MCP Client node, custom Python or TS agents). MIT licensed; source on GitHub.
  • Two install paths: the hosted endpoint at https://api.ejentum.com/mcp, or the stdio package via npx -y ejentum-mcp.
  • Backwards compatible: the MCP server is a thin wrapper over the harness API. No request/response schema changes.
  • Walkthrough: see the MCP integration guide for per-client install steps and the comparison vs HTTP Request and skill files.

v1.0. March 2026

Production launch.

  • 679 cognitive abilities across the four harnesses: reasoning (311), anti-deception (139), code (128), memory (101).
  • Six reasoning dimensions: Causal, Temporal, Spatial, Simulation, Abstraction, Metacognition.
  • Pre-rendered injection response: a procedure, a reasoning topology, a cognitive payload (Amplify / Suppress / Cognitive Style / Elasticity), a verification check, a named failure pattern, and a correct-shape example.
  • Rate limiting: 100 requests/minute per key.

Stability guarantee

The response schema is stable. Field names, section delimiters, and JSON structure change only additively, with advance notice.