These came out of overnight Nightshift runs. Each one opens the original file on meterless.ai.

A cistern, lit from above One HTML file. 200 attempts, 94 validated, attempt 148 kept. Open it ↗	The Atlas of Small Attractions A world to play in. Herd a living field with shifting rules. Play it ↗	After hours, and a scene scored Film plus music. 141 attempts, 3 local models. Watch · Listen ↗
Last light, and 29 more frames Stills from a story bible. 299 attempts, 1 local model, 5h 35m. Full size ↗	18 business deliverables Memos, plans and RFP answers, each checked against an answer key. Every draft is published. Read them ↗ · Replay a night in 3D ↗	50 textured 3D models From a sentence, a photo or a sketch. 246 of 246 pass the Khronos glTF validator. Turn them ↗

More in the gallery. The source records for these files are in provenance.json.

A prompt to one model returns one answer, and different models' first answers tend to look alike. Nightshift runs a job hundreds of times across several local models and groups the results into distinct approaches. In four real races, new approaches were still appearing after attempt 250.

Each night ends by condensing its strongest cluster into an insight pack. The next night starts from that pack instead of a blank prompt, and recipes that keep winning are saved as Methods and Skills you can read. No model is retrained.

How a night works · Every job mode, measured

Nightshift has a broad execution surface for a local AI application.

_{Recorded from the Nightshift field renderer on the How it works page. The page drives the field with a sample job, marked DEMO.}

1. The attempt. One prompt, one answer. This is how most people use agentic AI today.
2. The variance filter. One model runs across personas and temperatures, so no two attempts match.
3. The search. A lane per model. Every finished attempt is kept.
4. The scribes. Small readers summarize and tag each attempt, so the night becomes searchable.
5. The cluster. Related attempts gather into families. Closer to the core means more coherent, not proven.
6. Cluster synthesis. Members of a cluster are combined into a new, fuller answer.
7. Your AI hotspot. Any app that speaks the OpenAI API can send its requests to the same machine.

During the day the same machine is an AI hotspot. Any app that speaks the OpenAI API can use it: Cursor, Claude Code, the Meterless apps, and the experiments below. How the hotspot works.

Hardware notes for every edition are in editions.md.

Every number comes from a job log, a harness record or a test report, and names its machine. Every edition is now measured: Windows, Mac M5, Mac M1 to M4 and DGX Spark.

All benchmarks · Live job replays · Example output · What they taught us

Nightshift is the paid product. These are free, and each can use your Nightshift hotspot as its model provider.

Gaia A workspace that keeps projects, memory, decisions and artifacts across sessions and models. Watch it build an app.

Relay Describe a desktop job in one sentence. Relay plans it, runs it, checks each step, and replays it later with almost no model calls. Watch a run.

Swarms Type one goal and Swarms designs an agent team of 300+ agents in your browser. You keep the graph and can rerun it. Launch it.

Apps that only make sense when a model call costs electricity instead of a per-token fee. All experiments.

Meterless City Talk out loud to anyone in a photoreal city. About 700 NPC turns an active hour, $0 per conversation.

Sitcomify One photo in, a 90s sitcom pilot out. Dream Mode keeps writing and filming until you stop it.

DataPlant A customer-success platform that re-scores every account in the background. 0 cloud calls.

Everything above runs on four open context engines: memory, state, long-run reasoning and intent routing. Each one is an AGENTS.md implementation spec with a runnable reference and a conformance suite, so a coding agent can build it into your stack and prove it got it right.

The proof runs on your machine in 90 seconds with no API keys. The same 12-step task runs cold, then warm. The warm run starts with H-MEM memory and finishes in 8 chunks instead of 12. The output below is real; tokens are estimated at chars/4.

npx tsx examples/memory-compounding-research/index.ts

Clone one engine and hand it to your coding agent:

npx degit meterless/meterless/engines/hmem my-hmem
cd my-hmem/reference && npm install && npm test    # a working reference, green in about a minute

Your agent is done when the conformance suite passes: HMEM_IMPL=<your build> npx tsx conformance/runner.ts. Read the engines overview and the architecture.

Any AI that can read a link can read this repo. Paste this into ChatGPT, Claude, Cursor or Copilot:

Read the Meterless open-source repo at https://github.com/meterless/meterless.
Start with llms.txt and README.md, then AGENTS.md.
1. Tell me in five bullets what Nightshift and each engine do.
2. Ask me three questions about what I'm building.
3. Give me the exact steps for the one thing I should try first.

More in the prompt library.

From the benchmarks. The full list, with the run behind each lesson, is in learnings.md.

1. Count finished, accepted work. Tokens per second does not tell you that.
2. Use cloud models when someone is waiting. Use your own hardware for work nobody is waiting on.
3. Judging was the bottleneck. A dedicated judge model fixed it.
4. A generated app that loads may still not work. A check that could not run never counts as a pass.
5. Save what worked as written methods. No retraining is needed.

• Hackathons. The Open Agent Hackathon 2026 runs October 15 to 20, online, with up to $20,000 in prizes. Meterless has hosted 9 global hackathons with 30,000+ builders through GenAI.Works. Hackathon guide.
• Roadmap. Nightshift for Windows is out now. Nightshift for Mac ships Q4 2026. The next engine specs are swarm-orchestration and runtime. ROADMAP.md.
• Contributing. Start with CONTRIBUTING.md. Coding agents start at AGENTS.md; other agents at llms.txt.

Everything in this repo is Apache 2.0. App installers ship from each app's own repo and from meterless.ai.