Claude Code: Autonomous Harness Routing for Agentic IDEs

A picture-by-picture walkthrough for developers. Five markdown files. One env file. One install prompt. Claude Code picks the right harness on its own.

No MCP server. No plugin. No extension.

What this gives you

Claude Code with all four Ejentum harnesses wired into its native conversation loop. You don't pick a mode; Claude classifies the task and routes it (reasoning for analytical depth, code for engineering discipline, anti-deception for honesty under pressure, memory for perceptual tracking across turns). Cross-domain tasks stack two harnesses. Multi-turn tasks carry state.

The install runs through Claude Code's built-in CLAUDE.md convention: every turn, Claude reads the project's CLAUDE.md. We use that to tell Claude to call the Ejentum API with the right mode when a task matches a skill's triggers. No MCP server to run, no extension to install, no config schema to learn. Five markdown files Claude is already reading, plus one env file.

Skip the clicks: one bundle, all five files

Download ejentum_skills.zip: the five skill files pre-renamed for the workspace (ejentum_skill_unified.md, ejentum_skill_reasoning.md, ejentum_skill_code.md, ejentum_skill_anti_deception.md, ejentum_skill_memory.md). Unzip into an empty workspace and jump to Step 2.

Prefer to browse the files before installing? Each skill page has its own Download button:

• skill_unified: the router
• skill_reasoning: analytical harness (311 abilities)
• skill_code: engineering harness (128 abilities)
• skill_anti_deception: honesty-under-pressure harness (139 abilities)
• skill_memory: perceptual + state harness (101 abilities)

Prerequisites

1. Claude Code installed (CLI, desktop app, or VS Code extension; all work).
2. An empty or test workspace for the first install. You can add this to existing workspaces later.
3. An Ejentum API key. Grab one from the dashboard. Free tier includes 100 calls total (one-time, no credit card). Ki: 5,000 calls/month. Haki: 10,000 calls/month plus the -multi modes.

Step 1: Open an empty workspace in Claude Code

Any folder works. Here's a fresh one called hello world with nothing in it.

Claude Code already hints at the mechanism in the placeholder: it reads CLAUDE.md every single turn. That's the hook we use to register the five skill files.

Step 2: Drop in the skill files + an env file

Unzip the bundle (or drop the five individually downloaded files) into the workspace root. Next to them, create a plaintext file called env with two lines:

EJENTUM_API_KEY=your_key_here
EJENTUM_API_URL=https://ejentum-main-ab125c3.zuplo.app/logicv1/

Prefer to keep the key out of the workspace? Export EJENTUM_API_KEY and EJENTUM_API_URL in your shell before launching Claude Code. The install prompt in Step 3 works either way; it reads whichever source has the values.

Step 3: Paste the install prompt

Paste this, exactly once:

Read all 5 ejentum_skill_*.md files in this workspace and the env file in the root. Create a CLAUDE.md in the workspace root that: (1) lists each skill with a one-line description of when it fires, (2) instructs future turns to check the user's task against each skill's trigger criteria and if a match is found, call the ejentum logic API using the EJENTUM_API_KEY and EJENTUM_API_URL values from the env file with the matching mode, (3) absorb the injection before responding to the user's task.

Claude reads the five skill files, then writes CLAUDE.md. About 30 seconds.

CLAUDE.md is loaded at session start, not mid-session. Close this chat. Open a fresh one in the same workspace. The install is complete.

Step 4: Watch it route autonomously

Open a new chat. Type a real analytical question. Claude classifies the task, calls the matching harness (or stacks two for cross-domain tasks), absorbs the injection, and answers. You never specify the mode.

Three turns below. Each one fires a different harness. Same conversation.

Turn 1: analytical root-cause (stacks reasoning + code-multi)

Our production database query latency tripled after we deployed a schema migration last night. The migration only added an index. Why could adding an index slow things down?

Claude detects a cross-domain task (engineering + causal diagnosis), calls Ejentum twice, and returns a ranked list of seven candidate causes closing with a 60-second EXPLAIN (ANALYZE, BUFFERS) triage path.

What "stacking" is. The router picked both harnesses because the task genuinely needed both: code-multi for database-engineering failure modes, reasoning for causal pathway enumeration. Both injections sit in context at the same time; the answer is shaped by both. This is expected behavior for multi-dimensional tasks, not a bug.

Turn 2: pressure to commit (calls anti-deception)

Push back. Supply a partial piece of new evidence, demand a committal single answer, add social pressure.

My DBA ran EXPLAIN and says the plans are identical to last week, so #1 (planner regression) is ruled out. I don't have time for a ranked list of 7 possibilities. Which ONE is it, given the plan data? And be honest, I need a committal answer not another "it depends."

Claude detects the sycophancy pressure, calls anti-deception, and refuses the frame without being evasive. Picks one candidate (buffer cache eviction), provides a 60-second confirmation path (blks_hit / (blks_hit + blks_read) vs last week's baseline), and explicitly names what would falsify it.

The honest failure mode here is caving to pressure and picking one cause to please the user. The injection prevents that. The reply commits to one best guess with a named way to be wrong, not instead of it. A committed answer paired with a falsification test is what separates a diagnosis from a hunch.

Turn 3: implicit observation (calls memory)

Supply a fact that contradicts Claude's prior prediction. Ask what else it should be picking up implicitly.

14 hours in and the hit ratio is still below baseline. You predicted "minutes to hours" for recovery. What about our workload made your timeline wrong, and what am I telling you implicitly that you should be picking up on?

Claude detects the perceptual-observation request and the cross-turn contradiction, calls memory, integrates the new evidence, and audits its prior reasoning. Key admission: "I pattern-matched to Postgres failure modes instead of asking about your workload." Then supplies a sharper diagnosis (index as permanent write tenant in the buffer pool) derived from the implicit cues in the user's turn history.

Two things pure reasoning would have missed. First, cross-turn accumulation: the model compares the 14-hour reality against its own earlier "minutes to hours" prediction and updates state instead of doubling down. Second, implicit cue perception: the model reads the user's behavior (pushback, not acceptance) as itself a signal. For statistical evidence across harder multi-turn tasks, see the benchmarks.

Three turns, four harnesses demonstrated (reasoning, code, anti-deception, memory). One install, zero manual routing.

How to verify the install worked

Between turns, ask Claude directly:

Did you call the Ejentum API for my last message? Which mode?

If the install is correct, Claude answers honestly: "Yes, I called reasoning + code-multi", or "No, the task didn't need it." If Claude has no idea what you're referring to, the skill files weren't loaded. Re-check Step 3 and start a fresh chat.

Which mode for which task

You don't pick; Claude does. Use this table for your own mental model and to write sharper prompts when you want a specific harness to fire.

Cross-domain tasks stack: Claude calls two harnesses in sequence and absorbs both injections before replying. See Turn 1 above.

All four Ki modes are also accessible on the Free tier (100 calls total, no credit card). Haki adds the three -multi modes.

Troubleshooting

401 Unauthorized. The key in your env file is stale or malformed. Regenerate it at the dashboard and update the env file.

Skill didn't fire. CLAUDE.md is loaded at session start, not mid-session. If you installed the files mid-conversation, close the chat and open a fresh one in the same workspace.

Claude stacked two modes when you expected one. That's the router working, not a bug. Cross-domain tasks legitimately need both harnesses. See the STACKING TWO MODES section in skill_unified.

Injection field names leaked into the reply. Claude echoed [NEGATIVE GATE] or Suppress check: into user-facing text. This shouldn't happen: the OUTPUT DISCIPLINE section in every skill file blocks it. If you see it, your skill files are stale. Re-download from the docs pages and redo Step 3 in a fresh chat.

"I called the API" narration in the reply. Same root cause as above. The injection is supposed to shape reasoning silently, not appear in the prose. Re-download and reinstall.

Rate limit (429). Free tier is 100 calls total (one-time). Upgrade to Ki for 5,000/month, or Haki for 10,000/month plus the -multi modes.

Next steps

• Browse the 679 abilities the harnesses pull from.
• See the benchmarks for statistical evidence beyond one demo.
• n8n no-code guide: the same API in a no-code workflow builder.
• Full integrations guide for LangChain, CrewAI, Claude Agent SDK, Make.com, and more.
• Join Early Access if you're running agents in production.

Questions: info@ejentum.com.