demo-elasticity

Watch the demo

An AI agent that resolves GitHub issues in an OpenComputer sandbox with live resource scaling. The sandbox starts small, scales up for heavy compilation, and scales back down — paying for peak resources only when needed.

What it does

1. You comment @myagent resolve this on a GitHub issue
2. A webhook triggers the API, which spins up an OpenComputer sandbox
3. An AI agent (Claude) clones the repo, reads the issue, investigates, and makes a fix
4. The agent detects the sandbox has limited memory, scales up to 8 GB via the metadata service
5. Builds the Rust project, scales back down to baseline
6. Runs tests, commits, pushes a branch, opens a PR, and comments on the issue

The whole flow takes ~2 minutes and costs ~$0.30 in API calls.

Architecture

GitHub Issue → Webhook → API (Hono) → OpenComputer Sandbox
                                         ├── Agent (Claude Agent SDK)
                                         ├── scales 896 MB → 8 GB → 2 GB
                                         └── clone → fix → build → test → PR

Three components:

Component	What	Tech
ingest-rs/	Rust data pipeline — the project the agent fixes	axum, serde, tokio, reqwest
agent/	AI agent that resolves issues	@anthropic-ai/claude-agent-sdk
api/	Webhook handler that launches sandboxes	Hono, @opencomputer/sdk, @octokit

Setup

Prerequisites

• Node.js 20+
• An OpenComputer API key
• An Anthropic API key
• A GitHub PAT with repo scope
• ngrok or similar for webhook delivery

1. Install dependencies

cd agent && npm install
cd ../api && npm install

2. Configure environment

# agent/.env
ANTHROPIC_API_KEY=sk-ant-...
GITHUB_TOKEN=ghp_...
OPENCOMPUTER_API_KEY=osb_...
OPENCOMPUTER_API_URL=https://app.opencomputer.dev

# api/.env
OPENCOMPUTER_API_KEY=osb_...
OPENCOMPUTER_API_URL=https://app.opencomputer.dev
GITHUB_TOKEN=ghp_...
GITHUB_WEBHOOK_SECRET=your-webhook-secret
ANTHROPIC_API_KEY=sk-ant-...
CHECKPOINT_ID=           # filled in after deploy
PORT=3000

3. Deploy the agent snapshot

This creates an OpenComputer checkpoint with Rust, Node.js, gh CLI, Claude Code, and the agent code:

cd agent
npx tsx scripts/deploy-manual.ts

Copy the checkpoint ID from the output and set it as CHECKPOINT_ID in api/.env.

4. Configure GitHub webhook

On your repo, add a webhook:

• URL: https://<your-ngrok-url>/webhooks/github
• Content type: application/json
• Secret: same as GITHUB_WEBHOOK_SECRET
• Events: Issue comments only

5. Run

# Terminal 1 — API server
cd api && npm run dev

# Terminal 2 — expose to internet
ngrok http 3000

Then comment @myagent resolve this on an issue.

How the elasticity works

The sandbox starts with ~896 MB of memory. The agent:

1. Checks current resources: curl http://169.254.169.254/v1/limits
2. Scales up: curl -X POST http://169.254.169.254/v1/scale -d '{"memoryMB": 8192}'
3. Builds the Rust project (needs the extra memory for compilation)
4. Scales back down: curl -X POST http://169.254.169.254/v1/scale -d '{"memoryMB": 2048}'
5. Runs tests at lower memory (no recompilation needed)

The scaling happens live — no reboot, no checkpoint-restore. The VM's memory limit changes in-place via cgroup updates.

Redeploying

When you change agent code or prompt:

cd agent
npx tsx scripts/deploy-manual.ts
# Update CHECKPOINT_ID in api/.env with the new ID
# Restart api/

Testing locally

The agent can run outside a sandbox:

cd agent
npm run dev -- --repo diggerhq/demo-elasticity --issue 1

The elasticity API calls will 404 (no metadata service) but the build will work if your machine has enough RAM.