Newest signal 481h oldHow the evidence is collected →

CodeKudo
Attack my idea

Model Context Protocol

Your coding agent should know why the idea might fail

Connect Claude Code, Cursor or Codex — or any MCP client. Your agent can attack an idea before it writes a line of code, open the evidence behind any claim, and pull the build spec in its own format.

Why this matters more than it sounds

An agent handed an optimistic brief writes optimistic code. It widens the scope, builds the things on the “not yet” list, and never mentions the risk — because nothing in its context told it there was one.

Every response from these tools carries the verdict, the confidence, the scope exclusions and the kill criteria. The server even tells your agent, on connect, not to start building when the verdict is not “build candidate”.

The nine tools

search_ideasfree
Search published opportunities by segment, difficulty and time to MVP. Returns verdicts and confidence, not just titles.
get_ideafree
The full evaluation of one idea: who feels the problem, who pays, the concierge version, what NOT to build, risks, and kill criteria.
get_evidencefree
The source records behind an idea — supporting, falsifying, payment evidence and what is missing. Every item links to the original post.
get_specfree
The build spec in your platform's exact format: CLAUDE.md, .cursor/rules, AGENTS.md, a Lovable knowledge base, a Bolt prompt under its word ceiling.
get_usagefree
Plan, remaining credits, recent research runs and generated specs. Call it before spending.
research_startcosts credits
Start a deep research run and get a run id back. Long research does not fit in one request, so this returns immediately instead of holding the connection open.
research_statusfree
Where a run is: stage, waves completed, evidence gathered so far, and what it is still missing. Poll this.
research_resultfree
The finished verdict with its claims, counter-evidence and every source behind them.
research_upgradecosts credits
Move a finished run to a deeper tier when the remaining unknowns are worth more research.

Setup

  1. 1

    Create a token

    Go to Settings → MCP access and create one. It is shown once — we store only a hash, so we cannot show it again.

  2. 2

    Add the server to your client

    Endpoint: https://codekudo.com/api/mcp

    Run this in your project folder. It writes the server into the local config, so it applies to that project only.

    claude mcp add --transport http codekudo https://codekudo.com/api/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

    Check it with claude mcp get codekudo — it should say Connected.

    Using it inside VS Code

    The Claude Code extension runs the same CLI, so the command above is all you need — run it in VS Code’s integrated terminal (⌃`) from your project root. There is nothing extra to install on our side.

    Any other VS Code agent that speaks MCP — Continue, Cline — takes the JSON on the Cursor tab instead: same URL, same header.

    No Marketplace extension yet. Saying otherwise would send you looking for something that is not there.

    All three connect to the same standard HTTP MCP endpoint. Claude Code is the one we test end to end against every release. Create a token to get started.

  3. 3

    Ask your agent to attack something

    Try one of these:

    • Attack my idea with CodeKudo: an AI website builder for hairdressers in Germany.
    • Find CodeKudo ideas that are B2B, buildable in two weeks, and have payment evidence.
    • Show me the evidence behind that verdict — I want to read the original posts.
    • Pull the CodeKudo spec for that idea in Claude Code format and start on ticket one.

What the token can and cannot do

  • Scoped to your account. It reads published ideas, your own attacks and your own specs. It cannot see another member's private work.
  • Stored as a hash. We keep a SHA-256 digest, not the token. If our database leaked, the tokens in it could not be used.
  • Revocable instantly. Revoking takes effect on the next request, with no deploy.
  • It will not spend your credits silently. 2 of the 9 tools cost credits — research_start and research_upgrade. Each one states its cost in its own tool description, so your agent sees the price before it calls. Everything else reads.
  • Scoped, and read-only by default. A token you create can read the catalogue, analyze (run research — this is the one that spends credits) or export spec bundles. Only the boxes you tick are granted, and a tool the token cannot call is not even listed to your agent. Tokens expire after 90 days unless you say otherwise.
Create a tokenTry it in the browser first