I shipped ejentum-mcp today: four cognitive harnesses as MCP tools

#ai #mcp #claude #agents

Just shipped ejentum-mcp, an MCP server that exposes the four Ejentum cognitive harnesses as MCP tools any agentic client can call. One install, works in Claude Desktop, Cursor, Windsurf, Claude Code, n8n's MCP integration, and any other MCP-compatible client.

If you don't know Ejentum: it's a cognitive scaffolding API I've been building. The reasoning gap is structural, not informational. Models know plenty; they take shortcuts under pressure. The scaffold blocks the shortcuts.

You send a task description, you get back a structured cognitive scaffold (failure pattern to avoid, procedure, suppression vectors, falsification test) that the calling LLM absorbs internally before responding. The point is to catch LLM failure modes that ship to production as confidently-wrong answers: sycophancy under user pressure, hallucinated citations, causal shortcuts, reasoning decay across long chains.

Until today, integration meant either an HTTP request tool (in n8n or any framework that can POST), a skill file (for Claude Code's CLAUDE.md convention), or a direct Python/TypeScript call. All work, but each is bespoke.

The MCP server collapses that. One install captures the four harnesses (harness_reasoning, harness_code, harness_anti_deception, harness_memory) as native tools your agent can call.

Install

Easiest path is Smithery's one-click:

npx -y @smithery/cli install ejentum/ejentum-mcp --client claude

Replace claude with cursor, windsurf, cline, etc. Paste your EJENTUM_API_KEY when prompted. Done.

Manual install (any MCP client):

{
  "mcpServers": {
    "ejentum": {
      "command": "npx",
      "args": ["-y", "ejentum-mcp"],
      "env": { "EJENTUM_API_KEY": "your_key" }
    }
  }
}

Free tier: 100 calls, no card required.

The four tools

Tool	Use for
`harness_reasoning`	Multi-step analysis, planning, diagnostics
`harness_code`	Code generation, refactor, review, debugging
`harness_anti_deception`	Sycophancy pressure, hallucination risk, manipulation pressure
`harness_memory`	Perception sharpening, drift detection across turns

Each tool takes one argument (query, a 1-2 sentence task framing). Returns the harness scaffold as text. The calling LLM absorbs it internally and shapes its response with it.

The honest note on autonomous routing

This is the part most MCP server READMEs skip. I'm putting it up front because it's the truthful UX:

The tools fire reliably when you explicitly invoke them ("use the harness_anti_deception tool to evaluate..."). Soft suggestions also work ("reason about this", "check this for sycophancy", "review this code carefully").

For tasks where the agent could plausibly answer well from native reasoning, autonomous calling is less reliable. This is a property of optional MCP tools in general, not specific to ejentum-mcp. Agents are tuned to minimize unnecessary tool calls. Even with a thorough description rewrite (imperative "Call BEFORE answering", concrete trigger phrases, value props, DO NOT CALL exclusions), the v0.1.1 dogfood test showed the model still didn't fire on cold prompts.

For Claude Code users who want stronger autonomous routing, install the skill files alongside the MCP server. The skill files give Claude system-level context about when to call each harness. They coexist with the MCP install cleanly.

Why MCP for cognitive infrastructure

The most-installed MCP server on Smithery is Sequential Thinking. It exposes one tool that wraps one cognitive operation, and developers install it in droves. That's the demand signal: developers want callable cognitive operations as tools, with low friction and zero new accounts.

Ejentum has 679 engineered cognitive operations across four harnesses. The MCP server is the retail packaging that puts that library on the shelf where developers shop.