Mirako

lumiko

Docs an agent can actually use.

It scans the codebase, sends the source to Claude, and writes structured docs back. No custom parsers, no AST walking — Claude reads the code and writes what it found.

Current state Stopped · unfinished
Runs on Claude Code, your existing subscription. No API key.
Source mirako-studio/lumiko

Every agent starts by reading docs that stopped being true two refactors ago.

Nobody updates the architecture doc. The README describes an install path that changed last spring. Then an agent reads all of it as gospel and writes something confidently wrong, and you catch it in review if you catch it at all.

What lands in docs/

Four files. Three of them are for you.

your codebase lumiko scans Claude reads it docs/ written
For people
README.md

Project overview, install instructions, usage. The one that normally rots first.

For people
architecture.md

System design, module breakdown, Mermaid diagrams generated from what the code actually does rather than what someone drew in a meeting.

For people
api.md

Exported functions, types, signatures, examples.

For agents
context.json

The whole reason this tool exists. Not prose — a map.

The one that matters

Markdown makes an agent read three files and guess. This makes it look one thing up.

docs/context.jsonwritten for machines
{
  "project": {
    "name": "pricelantern-api",
    "summary": "Scrapes and ranks retailer prices",
    "language": "TypeScript"
  },
  "files": {
    "src/api/rate-limit.ts": {
      "purpose": "Token bucket limiter for the public API",
      "exports": ["consume", "reset", "RateLimitError"],
      "signatures": [
        "consume(key: string, cost?: number): Promise<boolean>",
        "reset(key: string): void"
      ]
    }
  }
}

Every file mapped to its purpose, exports and signatures. An agent that has this skips the search step entirely — it already knows which file owns the thing it was asked to change.

Getting it
# installs globally, uses the Claude Code CLI you already have
npm install -g lumiko

cd your-project
lumiko init       # writes .lumiko/config.yaml
lumiko generate   # writes docs/

Where it stopped

Being straight about this, because the install line above works and that can mislead.

What works

Small and medium repos

It runs end to end and the output is genuinely useful. The context.json is the part that earned its keep — it is what pointed at this whole idea of giving agents a surface to read.

Why it stopped

Not finished, and not pretending

Work moved to issues, which is the same idea aimed at a surface people update daily rather than one they regenerate. lumiko is stopped rather than done, and it is published in that state on purpose.

Still worth a look

If you want a context.json for your own repo, it will probably give you one. If it falls over on a large codebase, that is the known edge and not a surprise.

It is open, it is stopped, and it is not going to quietly become a product with a pricing page.

Picked it up?

If you have run it on something big and it held, we would genuinely like to know — that would change whether it stays stopped.

hi@mirako.computer