Ennodia. AI makes mistakes

But what one misses, another catches!

Paste this into your agent

try-ennodia.cherninlab.com
Read docs GitHub
Agents

Compatible local agents

Ennodia works through supported local harnesses. It does not provide a hosted model account or replace the subscriptions those tools use.

  • Codex
  • Claude Code
  • OpenCode
  • Kilo Code
  • Kiro
  • Cline
  • Hermes Agent
  • Antigravity
FAQ

Common questions

What is Ennodia?

Ennodia is a local Model Context Protocol (MCP) server for agent review. Your primary agent can request second opinions and compare the answers.

Why would one AI agent request help from other agents?

Different agents find different defects. Use additional model work when quality, risk, or plan uncertainty is important.

How long does a run take?

A run usually takes minutes. Ennodia starts local agent command-line interfaces (CLIs). Compare then adds a Judge task and a Result Advisor task.

Can Ennodia suggest a team for my task?

Yes. The Plan Advisor proposes explicit harness, model, and skill assignments. Ennodia validates the proposal as inert data. A separate call starts it.

Do I need to know Model Context Protocol (MCP)?

No. Paste the install line into Codex, Claude Code, Antigravity, or another agent. Tell the agent to install Ennodia.

Does Ennodia replace my main AI agent?

No. Your primary agent remains responsible. Ennodia routes a request to other local agents and returns their answers to the same workflow.

Which agents can Ennodia use?

Ennodia uses local harness adapters. Supported CLIs include Codex, Claude Code, OpenCode, Kilo Code, Kiro CLI, Cline, Hermes Agent, and Antigravity.

Does Ennodia expose an OpenAI-compatible application programming interface (API)?

The separate Ennodia IO package exposes experimental local provider options and a non-streaming chat-completions subset.

Does Ennodia send my code to Ennodia servers?

No. Ennodia runs locally and does not use an Ennodia-hosted service. A selected agent CLI can contact its configured model provider.

Why not manually copy prompts between tools?

Manual copies lose traceability and require you to compare the answers. Ennodia keeps agents, task IDs, failures, timing, findings, and results together.

Can Ennodia limit spend?

Ennodia estimates preflight input tokens. It can enforce maxChildTasks and maxEstimatedInputTokens. It cannot predict provider billing or internal harness context.

Is there a benchmark?

Yes. The documentation contains a dated bug-recall benchmark. Run bun run bench:bug-recall to repeat it from a checkout.

Does Ennodia support files and screenshots?

Available file and screenshot operations depend on each local harness. Provide file paths and short descriptions. Ennodia does not yet stage attachments.

Is Ennodia only for code review?

No. Ennodia also supports documentation audits, release checks, migrations, research checks, and design inspections.

Getting Started

Install Ennodia

Send the install link to your primary agent, or add Ennodia to an MCP client directly.

Agent prompt

try-ennodia.cherninlab.com

Command

npx -y ennodia

MCP config

{
  "mcpServers": {
    "ennodia": {
      "command": "npx",
      "args": ["-y", "ennodia"]
    }
  }
}
Use cases

Ennodia use cases

  • Persistent bug

    Run some agents when one answer is not sufficient. Compare likely causes before you edit.

  • Planning tradeoff

    Request migration plans from different models. Compare their assumptions before the project selects a plan.

  • Team suggestion

    Use the Plan Advisor to propose a bounded mix of harnesses, models, and skills. Inspect the plan before you start it.

  • Skill comparison

    Run one task with a skill such as source-grounded-audit. Run a second task without it. Compare the results.

  • Large plan review

    Divide a broad proposal into focused audit slices. Send the slices to different local agents. Compare the useful results.

  • Benchmark receipts

    Run deterministic bug-recall fixtures before you accept a claim. Repeat the test from the checkout.

Example prompt

Try an agent review

This asks Ennodia to run other agents, compare disagreements, and return one next action. 

Use Ennodia to ask the best available local agents:
"Review this change before I ship it. Find the highest-risk issue, docs drift, and install confusion. Compare disagreements with a model and give me the next action."
Process

How Ennodia runs

  1. 1

    Request

    Your primary agent sends one request.

    Agent Ennodia
  2. 2

    Route or advise

    Ennodia selects a route. The Plan Advisor can propose an inert plan.

    Ennodia Plan Advisor Plan
  3. 3

    Review

    Each worker completes its explicit slice.

    A B C Compare
  4. 4

    Judge and advise

    The Judge maps evidence. The Result Advisor recommends an answer.

    Judge Result Advisor Answer and trace
Interfaces

Core and MCP

  • Ennodia MCP

    Available

    Agents use this shipped interface. It provides MCP tools for discovery, plan advice, runs, budgets, tasks, and Compare.

  • Ennodia Core

    Available

    Ennodia Core contains shared behavior for MCP and the IO package. It controls routes, advice, tasks, failures, and Compare.

Budget

Spend extra model work on purpose

  • Ennodia uses your local agent command-line interface (CLI) programs, subscriptions, and model choices as a review panel
  • Ennodia finds supported harnesses automatically and keeps adapters thin
  • The Plan Advisor proposes a bounded team as inert data. A separate validated call starts the plan
  • Ennodia estimates preflight input tokens and applies hard local limits
  • Ennodia tracks task IDs, time estimates, output, failures, Compare state, and terminal run history
  • The Judge maps parallel answers. The Result Advisor recommends an answer
  • Model and skill preference memory remains on the roadmap
Judge and Result Advisor

Clear roles before and after work

  1. Plan Advisor

    Before work, it proposes a bounded team as inert data. It never starts the plan. The caller must start it explicitly.

  2. Judge

    During Compare, it identifies agreements, contradictions, missing context, unique value, and risk.

  3. Result Advisor

    It combines the Judge findings into a recommended answer. If the Judge fails, Ennodia shows a candidates-only basis.

Receipts

Bug-recall benchmark

The current deterministic snapshot contains four TypeScript review fixtures. Ennodia parallel Compare scored 100% recall and 100% precision in fixture mode. The benchmark page lists its limits.

Read the benchmark snapshot

Credits

Inspired by karpathy/llm-council

That project demonstrates this pattern. It requests independent answers from some models. It then compares and combines the answers. Ennodia applies the pattern to local agent workflows, MCP, task traces, failure visibility, and harness adapters.

Get started

Request second opinions

Paste the install line into your primary agent. Tell it to use other agents when one model is not sufficient.

Install GitHub