Your next incident deserves a coworker.
One that finds the bug, writes the fix, and brings the evidence.
Uptime Harness connects error reports to an AI repair loop: reproduce, repair, test, open a pull request, and verify the deployed result. Every step stays on one ticket.
Sentry sends the error and its stack trace.
Codex works in isolation. Tests keep it honest.
A pull request, a verified release, a clear timeline.
Use failed checks and trace findings to guide the retry.
Docker runs the dashboard, Service Desk and repair controller. Bring your own MongoDB Atlas, OpenAI, GitHub and Sentry accounts.
Install Git, Docker with Compose v2, and Make. Docker supplies the application runtimes.
git clone https://github.com/aemal/uptime-harness.git
git clone https://github.com/aemal/uptime-harness-service-desk-demo.git
cd uptime-harness
cp .env.example .env.localFor your own repairs, fork the Service Desk and clone your fork as the sibling folder.
.env.local.Add your Atlas connection, OpenAI key, GitHub token and Sentry DSN. Generate the two local secrets below.
# Paste into OWNER_TOKEN
openssl rand -hex 24
# Paste into SETTINGS_ENCRYPTION_KEY
openssl rand -hex 32Set REPAIR_REPOSITORY to your fork and choose a CODEX_MODEL available to your account. Keep the environment file private.
Then create the Atlas indexes and telemetry collection, and check the repair worker's configuration.
make up
make setup-atlas
make doctorUnlock with OWNER_TOKEN. Atlas stays cloud-hosted; a healthy container does not prove provider connectivity.
SENTRY_DSN before starting. With Cloudflare's CLI installed, run cloudflared tunnel --url http://127.0.0.1:3028 in another terminal. Save its HTTPS URL in Project settings, configure https://YOUR-TUNNEL/api/webhooks/sentry on a Sentry internal integration, save the signing secret in Uptime, and add the integration to an issue-alert rule. Keep the tunnel running.service-desk after a repair. For automatic deployment and verification, connect the fork to Vercel and set DEMO_PRODUCTION_URL to your production URL. Leave it blank until that deployment is configured.make up again after environment changes. The demo reset button is currently restricted to the maintainer's demo repo; forks need an explicit allowlist change.The trusted repair controller mounts the Docker socket to launch isolated workers. Run it on a machine you control.
Full setup, credentials and deployment guide ↗The harness owns the handoffs. You get a durable repair board, clear evidence, and control over what ships.
Protected checks stay outside the agent’s editable workspace.Swipe the diagram to explore →
Atlas database, Search, GridFS and time-series telemetry were exercised in the live repair. Change streams power live updates with a polling fallback. Vector Search and Voyage AI are implemented as optional repair memory and were not exercised in that run. Atlas Agent Engine is a future integration. Raindrop diagnostics can guide a failed attempt’s retry; the recorded live repair passed on its first attempt.
Read the verification record ↗A signed Sentry webhook creates a Kanban ticket with the issue, error details, and source links. MongoDB keeps the history.
Codex reproduces the failure and works in an isolated checkout. Protected regression checks sit outside the code it can change.
Passing checks gate the pull request. Optional auto-merge is followed by a deployment check before the incident is marked resolved.
Raindrop traces and test failures inform the next repair attempt. Verified repair memory helps later investigations; the tests stay protected.
The ElevenLabs incident coworker can explain the latest evidence in a web call. Optional Twilio calling keeps the person on duty in the loop.
Watch an AI service desk hit a real tool error, follow the repair through GitHub, then retry the same request after deployment.
More builds on AgentGeeks ↗A growing collection of repair targets. Separate from the harness, so each fix has its own pull request and deployment.
Describe an IT issue. An AI coworker assesses the impact, recommends next steps, and calls a tool to create a support ticket.
A simple contact form that started it all. Add a contact, leave optional notes blank, and see the saved result.
Use sample information in these shared demo apps. View all examples ↗
Read the code, run the stack, adapt the repair loop to your team. All three repositories are MIT licensed.
Dashboard, repair worker, provider integrations, Docker setup, and protected checks.
AI Service Desk, contact manager, regression checks, and the end-to-end demo guide.
A small static site. No database, no build step, and no API keys.
Run the harness on a Docker-capable machine. The repair worker needs a persistent process. The example apps and this landing page can run on Vercel. Bring your own provider accounts and follow the repository setup guides.
Start here ↗Built at the Harness Engineering & Model Wrangling Hackathon at SHACK15, San Francisco. Thank you to the sponsors, organizers and everyone who shared their time and ideas.
Atlas keeps the repair history, searchable incidents, evidence files and telemetry together. Its optional vector memory gives the next investigation a place to look back.
Database · Search · GridFS · change streams · time seriesEvery repair attempt can carry a trace of the commands, failures and test results. Diagnostic signals help the harness give the next attempt more useful context.
Attempt traces · failure diagnosis · retry guidance
Samantha is the voice of the harness. Web conversations use the latest ticket evidence, with optional Twilio calling for the person on duty.
Samantha · web voice · phone integrationAnd to Cerebral Valley: thank you for bringing the builders together. To the mentors, judges and SHACK15 team, thank you for the space, feedback and energy.
The hackathon ↗
OpenAI
Sentry
GitHub
Vercel
DockerTwilio phone callsOpenAI generates the repair and powers the demo’s triage. Sentry reports errors; GitHub hosts code and checks; Vercel deploys the examples; Docker isolates the worker.
A handful of snapshots from a day of building, exchanging ideas and watching the bay.
I create AI personas tailored to your team’s roles and workflows, then share the knowledge through a curated, hands-on AI workshop. Your team learns how to use and adapt them independently.
For engineering, product, management, marketing, revenue, growth, outreach, and GTM teams. Each workshop is shaped around your goals, with practical lessons in context engineering, Agent Harnessing, and Harness Engineering.