From one possibility to a body of work you can inspect, keep, and continue.
01 / THE ATTEMPT
How you use Agentic AI today.
You prompt an agent, read the answer, adjust, prompt again. Every round is one attempt at the outcome you asked for. But what if you could explore the entire set of responses the model could give instead of going with the first viable one you find.
One light in the field. One finished attempt.
02 / THE VARIANCE FILTER
Turn one model into many.
Point Nightshift at any model, local or cloud. A variance filter runs it across different personas and temperatures, so no two attempts are the same.
Each glowing point is a lane, labeled with the model it runs. Its trails are attempts.
03 / THE SEARCH
Don’t stop at the first answer.
Let complementary models explore different approaches: a lane per model, local ones like Qwen, GPT-OSS and DeepSeek on your machine, cloud ones only if you add a key. Attempts accumulate while you’re away; a model too slow for a conversation has all night here. Every finished attempt is kept, so a useful idea doesn’t disappear just because it wasn’t first.
More perspectives from the capacity you already own.
04 / THE SCRIBES
A night that reads its own work.
Watch the small bone-white scribes leave the cluster core to read attempts, then return. They read, summarize, and tag what was made, making a whole night searchable. Judges and verifiers then help separate promising work from work that meets your bar.
Small diamonds. Short traces. A reading crew at work.
05 / THE CLUSTER
Find the signal in the possibilities.
Related attempts gather into clusters: families of similar answers. Now you can see where ideas converge, where they disagree, and which directions deserve a closer look.
Closer to the core means greater coherence—not proof.
06 / CLUSTER SYNTHESIS
Many attempts. A fuller answer.
A cluster holds different takes on the same idea. Gather its members, choose a direction, and synthesize a new answer from their useful details and different perspectives. Build on what the group discovered.
Gather members→Synthesize→Review
A new synthesis, kept beside the attempts it came from.
Everything in 01–06 isn’t locked inside Nightshift. Add Meterless as a provider in Cursor, Claude Code or any app that speaks the OpenAI API, and its requests land on your machine. Any computer running Nightshift becomes an AI hotspot that can power your apps.
Any AI app→Meterless provider→Nightshift on your machine
One machine, many apps, all drawing on the same night. The hotspot for your apps runs on the Mac, Windows and DGX editions.
A finished night is not a scrollback. Nightshift keeps every attempt, cluster, judgment and receipt. Continue run folds last night’s consensus, open disagreements and best rendition into tonight’s attempts. Compounds join finished nights into one job. What worked is kept as a Method or a versioned Skill.
NIGHT 1Blank brief202 attempts · 2 clusters
continue
NIGHT 2Starts from 2 clustersconsensus kept · gaps reopened
continue
TONIGHTBuilds on both1 accepted · receipts kept
Methods · Skillsevidence rises only as nights test them
↺ Tomorrow’s brief begins where tonight settles
The loop closes: brief, attempts, clusters, judgment, settled, and on to the next night.