TokenEater

A multi-agent pipeline where free LLMs work as a team — and check each other's work.

---

One LLM call on a complex request fails in unpredictable ways — and free-tier models fail more often than most. TokenEater splits the problem instead of retrying blindly: several agents cooperate as a team, and every answer passes verification before it ships.

How it works

request ──> GATEKEEPER ──> BRAIN ──> agents ──> final assembly
             │               │          │
        route or answer   plan &     research / code /
        directly          split      creative / critique

1. GATEKEEPER — classifies the request: answer directly, delegate to planning, or ask for clarification.
2. BRAIN — deep planning: breaks the task into concrete microtasks.
3. Agents — execute microtasks, each drawing a model from a pool of free OpenRouter models (research, code, creativity, criticism).
4. Final assembly — merges results, verifies quality against explicit criteria, feeds critique back to the worker until the answer passes.

Features

• Runs entirely on free OpenRouter models (:free tier)
• Automatic verification & correction: workers get critiqued and retry until they pass
• Strict JSON contracts between agents, parsed defensively (markdown-fenced JSON included)
• Handles complex multi-step requests that break a single model call
• Model pool is a plain dict in m.py — swap models without touching logic

Quick start

git clone https://github.com/alesrg/TokenEater.git
cd TokenEater
pip install -r requirements.txt

Put your OpenRouter API key into apiKey at the top of m.py, adjust the models dict if you like, then:

python m.py

Status

An experimental sandbox for orchestration ideas — the contract-and-verify pattern you'd use between any two services, applied to agents.

---

Part of alesrg's experiments · built in public