The London AI Cost Hackathon

Set up the sponsor accounts

You can skip this section until you want to make live model calls. If the event gives you credentials, use those instead of creating separate ones.

Merge Gateway

1. Open gateway.merge.dev/signup.
2. Create an account and complete verification.
3. Accept the event organization invitation if you receive one.
4. Open Settings → API keys.
5. Create a regular model-calling key beginning with mg_.
6. Copy the event project ID. Do not use an mgmt_ management key.

condense.chat

Start at condense.chat. There are two supported ways to use it, and you may use either or both.

Option 1: Run a coding agent through Condense

Install and sign in to the dense CLI by following the agent setup instructions. Then launch Codex or Claude Code through Condense:

dense codex
dense claude

This keeps the agent's normal authentication and routes its requests through the Condense proxy.

Option 2: Call models through the proxy API

Point an OpenAI-compatible client at Condense, supply your normal upstream provider key, and add your Condense token:

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["OPENAI_API_KEY"],
    base_url="https://api.condense.chat/openai/v1",
    default_headers={
        "X-Condense-Auth-Token": os.environ["CONDENSE_AUTH_TOKEN"],
    },
)
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Review this change."}],
)

The standard OpenAI request methods continue to work. Anthropic users can follow the equivalent proxy example in the Condense quickstart.

Store credentials locally

cp .env.example .env
chmod 600 .env

Edit .env:

MERGE_GATEWAY_API_KEY=mg_replace_me
MERGE_GATEWAY_PROJECT_ID=replace_me
CONDENSE_AUTH_TOKEN=ak_replace_me
OPENAI_API_KEY=sk_replace_me

Install the prerequisites

You need:

• Git
• Python 3.11 or newer
• uv

Check them:

git --version
python3 --version
uv --version

Fork and clone the starter

Fork the starter repository on GitHub and keep your fork public. Then clone your fork, replacing YOUR-USERNAME with your GitHub username:

git clone https://github.com/YOUR-USERNAME/ai-cost-hack.git
cd ai-cost-hack
uv sync

This installs the benchmark CLI and development tools from the committed lockfile.

Run the public benchmark

uv run costhack benchmark --public

The deliberately incomplete starter should report:

NOT ELIGIBLE  quality=50.0
Model cost is measured separately in Merge Gateway.

The starter catches five of ten cases. The examples demonstrate the input, output, and scoring contract without revealing the final test set.

Inspect a public case

uv run costhack inspect public-auth-delete

Read data/public_cases.json to see the complete public case schema. Public cases include their rubrics; private cases do not.

Edit your strategy

Open submission/strategy.py. Keep the review(case) entry point and return this shape:

{
    "risk": "high",
    "findings": [
        {
            "category": "authorization",
            "severity": "high",
            "file": "api/delete_project.py",
            "evidence": "The project is loaded without an owner check.",
            "explanation": "A user can delete another tenant's project.",
        }
    ],
    "tests": ["Cross-tenant deletion must return 403"],
    "next_action": "block",
}

If you use a model, create a Merge Gateway client inside your submission and read its credentials from the environment.

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MERGE_GATEWAY_API_KEY"],
    base_url="https://api-gateway.merge.dev/v1/openai",
)
response = client.chat.completions.create(
    model="gpt-5.5",
    messages=[
        {"role": "system", "content": "Review this change."},
        {"role": "user", "content": case["brief"]},
    ],
    extra_body={"project_id": os.environ["MERGE_GATEWAY_PROJECT_ID"]},
)

There are two complete starting points: submission/strategy.py contains the deterministic Python rules baseline, and examples/merge_only/strategy.py contains the GPT-5.5-through-Merge strategy.

Once quality is reliable, consider condense.chat for long, noisy cases. Use the dense CLI with Codex or Claude Code, or call models through Condense's supported proxy API. Scored model inference must still appear in Merge Gateway so its cost is included in the official measurement.

Validate your submission

uv run costhack preflight
uv run costhack benchmark --public
uv run pytest
git status --short

Confirm that .env is ignored and that your public benchmark still passes after every strategy change.

Commit, push, and submit

git add submission pyproject.toml uv.lock
git commit -m "Submit PatchGuard strategy"
git push

Submit your public GitHub repository URL through the submission form before the cutoff.