Give your computer a night shift.

02.1What each edition has

Shown as read from each edition's release notes and build files in October 2026. Where a document says "not confirmed", the feature has not been tested on that hardware.

02.2Minimum hardware

03.1First run: the three "Get started" steps

Setup shows a Get started card with three steps and a counter. Do them in order. The counter reads 3/3 when you are ready for a night.

1. Add a provider key or start a local model.

   Press Open Setup. Either paste a cloud key under Providers, or download a fleet under Fleets. See Models and providers.

2. Send a prompt, or point one app at the endpoint.

   Press Open Queue. Type a task and press Begin the night. Or copy the endpoint from Providers into your editor. See Connect your apps.

3. See your first receipt.

   Press Open receipts. Every answer carries a receipt: which lane won, the tokens, the cost and the cache status.

03.2Where your data lives

04.1What each screen is for

Find the right Setup tab

Open Setup in the top bar. Start with the tab that matches what you want to change.

Enlarge ↗

Console is the plug icon in the top bar (Ctrl+5 on Windows, ⌘5 on Mac). It shows connection settings, apps, activity and proof.

04.2Keyboard shortcuts

On Windows and Linux, use Ctrl where this table shows ⌘.

04.3Hints and Simple view

DGXShort hints can appear next to controls as you use the cockpit. They are off by default. Turn them on, down or off in SetupPreferencesHints. Start over resets which hints you have seen.

◂ Simple view in the top bar switches to a one-box view. Use it when you only want to watch traffic flow through the endpoint and prove it works.

05.1Start the local runtime

The runtime is the process that loads GGUF models. On the desktop editions it starts with the app. Check it here:

1. Open the runtime card.

   Go to SetupLocal models. Find the card Meterless Local Runtime. In the Web edition the tab is named Fleets.

2. Read the status pills.

   They show: running or stopped · meterlessd available · the number of GGUF models · backend · RAM · VRAM.

3. Start it if it is stopped.

   Press Start runtime. Use Detect hardware if the backend or VRAM reads unknown.

4. Choose where models go (optional).

   Press Choose models folder… to use another disk. Press Use default folder to go back. Open models folder shows the files.

sudo systemctl start meterlessd
systemctl status meterlessd

05.2Download a local model

The fastest way is a fleet: it downloads every model a crew needs (see Fleets). To add one model by hand:

1. Open the catalog.

   In SetupLocal models, scroll to Runtime model catalog · grouped by class.

2. Pick a class.

   Classes: C · Heavyweight, B · Workhorse, A · Fast & light, Embedding · Scribes, Companion · Feeders, and Media (Image, Video, Speech, Music, Transcription, 3D).

3. Press Download optimized.

   Download optimized picks the best quantization for your hardware. Download takes the listed file. A row marked Too large does not fit your memory.

4. Wait for "ready".

   The row goes through downloading N%, then verifying checksum…, then ready. Companion files (VAEs, encoders, vocoders) download with the model that needs them.

5. Seat it on a lane.

   A model on disk is not a lane yet. Press Start to use, or add it in the Lanes panel. See Lanes.

05.3Add a cloud provider key

1. Open Providers.

   Go to SetupProviders. Scroll to Cloud provider keys. The card reads "held by meterlessd · write-only · never logged".

2. Get a key.

   Press Get key ↗ on the provider's card. Create a key on the provider's site. Copy it.

3. Paste and save.

   Press Add. Paste the key. Press Save. The daemon checks the key with a cheap request that spends no generation tokens.

4. Read the verdict.

   A working key reads configured. Other verdicts: invalid_key, insufficient_credit, forbidden, rate_limited (this one proves the key works), provider_error, unreachable.

05.4The privacy floor

Every job and every connected app has a floor. The floor decides which lanes may answer.

Set the global floor in SetupPreferencesPrivacy floor. The preset Most private sets it to local-only in one click.

06.1The duties

Each lane has one or more duties. The Lanes panel groups lanes into duty rows. You can edit Race, Vision, Judge, Verify and Prep. Create and Retrieve follow from what the lane runs.

06.2Add a lane

1. Open Lanes.

   Press Lanes on the rail, or ⌘6. Stay on the Lanes & memory sub-tab.

2. Press + Add lane.

   The Add a lane sheet opens.

3. Pick the job.

   Choose a family: Language, Vision, Images, Video, Audio, 3D or Retrieval. Then a task, for example "Build & debug code", "Score different answers", "Generate an image" or "Search my information".

4. Pick the model.

   Models are ranked by what fits the free memory. Press 🎲 Autofit for the largest model on disk that fits. For a cloud lane press 🎲 Roll a cloud model.

5. Wait for ready.

   The card shows the state. If it reads insufficient_budget or low_live_memory, the lane did not fit. See Troubleshooting.

06.3Lane card actions

Each card shows one highlighted next action. The rest sit under ⋯ All controls.

06.4How memory is budgeted

Two gates run when a lane starts. Both must pass.

Open the ledger from the sticky left column of Lanes & memory: Open memory ledger. Mac M5+ A fleet larger than memory takes turns loading its models. A model that can never fit is greyed with Refactor fleet to fit.

07.1Pick a fleet from the catalog

1. Open the atlas.

   Go to LanesFleets. Switch between Atlas (a map by diversity and measured speed) and List. Featured fleets also appear at the top of SetupLocal models.

2. Filter.

   Use the chips. Needs download shows fleets with missing models. Browse by job: Site Builder, Show Runner, Research Review, Batch Converter, Campaign Studio, Security Review, Writers' Room, Knowledge Desk, Agent Desk, Night Studio, Build Shop, Document Desk, Visual Review, BI Desk, Tagging Swarm, Heavy Anchor, and more.

3. Open the card.

   Read the lane roster, the memory fit and the status line.

4. Activate.

   Press Download & activate if models are missing, or Activate if they are on disk. Lanes that changed stop and the new set cold-starts. If a switch is in progress you may see Switch anyway or Wait.

07.2Sizes and claims

A fleet's claim to code is graded: strong, capable or assist. A card says it can code only when it seats a strong or capable coder. A model enters the local registry only when the pinned runtime can load it.

07.3Build your own fleet

1. Start from what runs now.

   In the Fleets toolbar press Save current lanes. This snapshots the lanes running right now as a custom fleet.

2. Or start empty.

   Press + New fleet. The editor opens. Press + Add lane for each seat and pick a duty and a model.

3. Keep a judge.

   Give every night fleet one non-racing Judge lane. Do not seat a tiny or "thinking" model as judge.

4. Save.

   Press Save fleet. Custom fleets get Duplicate, Edit and Delete. Reset to definition removes manual lane edits from the active fleet. Reconcile starts lanes that are stopped.

07.4Three things a fleet is not

08.1Queue your first night, step by step

1. Open Queue.

   Press Queue on the rail. Leave the switch on Auto for your first night.

2. Write the task.

   Describe the work in plain words. Say what good looks like. Example: "Write a go-to-market memo for Ledgerly. Decision-ready, 900–1,300 words, with the arithmetic shown."

3. Attach files (optional).

   Drop a PDF, DOCX, spreadsheet, image or archive on the box. Every lane can read it. A scanned PDF with no text says so and goes to the document reader.

4. Read the plan.

   Under the box, Tonight's plan lists kind, attempt ceiling, diversity, budget, judge, window, residency and outcome. Auto labels each value: measured, estimated, default, carried, or your override.

5. Take the fleet or change it.

   The composer proposes a fleet. Press Use this fleet, or open Lanes and activate another.

6. Begin.

   Press ☾ Begin the night. Under quiet, the night waits for 22:00 and ten idle minutes. Press ⚡ Begin right now to start at once. On the Jobs board a waiting job has ⚡ Don't wait · run it now.

08.2The job kinds

08.3When it runs

08.4The Jobs board

Press Jobs on the rail (⇧⌘2). The list is on the left, the inspector on the right.

Schedules and watchers hide behind the {n} scheduled / watched toggle. Search the board by prompt, id, kind or judge.

09.1A night that keeps itself running

09.2Two postures: Swarm and Depth

The Monitor panel shows the posture: "swarm · full fleet" or "depth · new jobs wait for the quiet window". The change is written on the job's own timeline.

10.1The outcome word

The floors: policy honoured · verifier verdict · artifact valid · tests pass · bar met · completed · proof. The outcome is a projection over the record. It never re-judges and never guesses.

10.2Read the morning

1. Open Morning.

   On Night, press the ☼ Morning sub-tab, or ☼ Morning review on a finished night. Panes: Exposure, Details, Suggestions.

2. Read the answer.

   Press 📖 Read the answer for the winner in full. Straight to the winning answer does the same from the Archive.

3. Read the night report.

   It reads across all clusters: the answer-space map, agreement and disagreement, minority positions, judging, unresolved questions, the answer to take forward and the next experiment. Each listed cluster links to its inspector.

4. Watch the reel.

   The highlight reel cuts the night into up to sixteen moments from the event ledger. It makes no model calls. Press Watch the night as a short reel.

5. Export.

   Press ⇩ Export to download the night as a zip. It asks before it includes your prompt.

10.3The Library

Every answer as a card. Eight view modes: small, medium, large, review, streams, timeline, storyboard and repair. Sort, compare and pick.

10.4Clusters, reports and cluster chat

10.5Continue, refine, compound

10.6The Archive and semantic zoom

The Archive shows every run as a light. Rank by yield, attempts, duration, clusters, score or recency. Filter by status, time and judge mode.

Zoom out of a night (press Pull back or scroll out past the edge) and its attempts fold into cluster cores, then into the job's Archive dot. Zoom in on a dot and it blooms back into the night. Press Watch this night again in the field to replay it.

11.1The four judge modes

Configure judges under WorkersJudges. The judge deadline (60 s by default) is in Preferences. A judge that runs out of time abstains. It never guesses.

11.2Validators and the sandbox

11.3The Repair Queue

1. Open the Repair view.

   In the Library, choose the repair view mode. It gathers every answer that needs fixing. On DGX the ns repair list command shows the same queue.

2. Read the issues.

   Each recorded failure (validator, sandbox, test, judge or make evidence) becomes a typed issue with the cheapest action: recheck, repair or rerun. This diagnosis makes no model call.

3. Read the plan.

   The plan is shown under a hash before anything runs. One repair per lineage at a time, under a finite allowance.

4. Run it, or let free time run it.

   In SetupPreferencesRepair in free time set mode (mend · collect only · off), route (local lanes only · any lane) and promotion (when every check passes · I choose).

5. Read the result class.

   accepted repair · improved but not accepted · no change · regressed · unresolved · execution failed. Only a verified repair is promoted. A "done" row is never "accepted".

11.4Receipts and trace

12.1Find the endpoint and the key

1. Open the endpoint card.

   Go to SetupProviders. The Runtime endpoint card is at the top. The Console (⌘5) shows the same card.

2. Copy the values.

   Base URL http://127.0.0.1:8088/v1. Press Reveal then Copy on the auth token. Default model meterless/auto.

3. Or copy a recipe.

   Press Copy setup. Ten recipes are ready with the live URL and token filled in: curl, OpenAI SDK for Python and Node, Continue, Cursor, Cline/Roo, Zed, Open WebUI, LibreChat, and an experimental Anthropic SDK recipe.

4. Test it.

   Run the curl command below with your token. A JSON answer means the endpoint works.

curl http://127.0.0.1:8088/v1/chat/completions \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"meterless/auto","messages":[{"role":"user","content":"Say hello in five words."}]}'

# Python, with the official OpenAI SDK
from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8088/v1", api_key="YOUR_TOKEN")
r = client.chat.completions.create(
    model="meterless/auto",
    messages=[{"role": "user", "content": "Summarize today's PRs in 3 lines."}],
)
print(r.choices[0].message.content)

12.2Modes: the model names your app sends

Treat Nightshift as a provider whose model IDs are modes. You can also send a real model id, for example gpt-5.5 or local-runtime/llama-3.1-8b-instruct, and routing pins to it.

GET /v1/models lists the modes, your configured cloud models and installed local models. Chat cannot carry images or speech: asking for one answers 400 class_carrier_unsupported and points at the jobs route. Supported: streaming, tool calling, structured outputs, embeddings, /healthz.

12.3What a connected app gets

12.4Hotspot mode: serve other devices on your network

1. Turn it on, on the box.

   Go to SetupPreferences. The first card is Hotspot mode. It is off by default. Switch it on. The switch only works from the computer that runs Nightshift. A request from the network to flip it gets 403.

2. Open the firewall.

   Allow TCP port 8088 on this computer's firewall, on your private network only.

3. Copy a LAN address.

   The endpoint card now lists Base URL · hotspot rows, for example http://192.168.0.55:8088/v1. Use the same token.

4. Point the other device at it.

   Press Show QR. A phone camera opens a page on the box itself with the base URL, the key and paste-ready snippets. The key rides after the #, so it reaches no log.

12.5Pair your phone

1. Open Controls.

   Go to SetupControlsPhone. Switch it on. Pairing runs on port 8089 while the cockpit is open, LAN only.

2. Scan the QR.

   Scan it from a phone on the same network, or type the code RT-XXXX-XXXX. The link opens the live field, not a status card. Nothing is installed on the phone.

3. Choose what the phone may do.

   Scopes: status.read · jobs.read · queue.submit · job.write (pause, resume, cancel, review scribes) · lane.write (force start, enable, disable, recolour).

4. Rotate when needed.

   Rotate code invalidates the code for new pairs. Phones already paired stay until you revoke them.

12.6The ns command line DGXMac M5+Windows

ns talks to the daemon on this machine with the daemon's own token. It is read-mostly, loopback-only, and never touches lanes, fleets, models, keys or setpoints. Use it from scripts, CI and coding agents. On Windows it is %LOCALAPPDATA%\Nightshift\cli\ns.cmd.

ns discover                               # what this box can do right now
ns status --json                          # is the daemon reachable
ns route --prompt "refactor this module"  # how it would route, no model call
ns generate -f brief.md --n 64 --privacy local-only --wait
ns job watch job_abc                      # live NDJSON of the night
ns job outcome job_abc                    # Accepted / Not accepted / Not yet decided
ns job report job_abc --format markdown   # the night read across clusters
ns context get job_abc --view winner --format markdown
ns verify status job_abc                  # which generated apps ran in the sandbox
ns repair list job_abc --scan             # what failed and what could fix it
ns import preview ~/chatgpt-export.zip
ns proposals list                         # the inbox of suggestions

Exit codes: 0 ok · 2 usage · 4 not found · 5 conflict · 6 unavailable on this box · 7 saturated · 8 unauthorized · 9 invalid · 12 daemon unreachable · 13 deadline · 20 partial. A retried generate creates a second job. --window quiet waits until 22:00.

12.7Tools the models can use (MCP)

Nightshift is an MCP client. Under WorkersMCP, 13 servers install in one click: memory, filesystem, github, sequential-thinking, fetch, time, git, playwright, sqlite, postgres, brave-search, slack, everything. The daemon owns them, so they keep working with the cockpit closed. Twelve native tools ship beside them: web search, browser reading, file access, HTTP fetch, cluster context and conversion. Tools run only inside a job.

Nightshift also consumes runtimes you already run. Ollama, LM Studio, llama.cpp, vLLM and Forge are probed and seated as lanes.

13.1Install the media engines

13.2Make something from an answer

1. Open an answer.

   In the Library or the field, click an attempt. Press Make.

2. Pick a transform.

   The MAKE menu has 11 transforms: a document, a deck, a page, a standalone web app, a still, a clip, animation, narration, music, a transcript, or a conversion.

3. Wait for the lane.

   A make leases an engine under its own token. Heavy image, video and music cores take turns. A make waiting on a busy lane starts the moment a fitting lane is free.

4. Find it.

   Makes appear beside the answer they came from. The field tracks makes apart from attempts. The Making pill on the Night bar lists what is being built.

13.3Cloud media models

Image, video, speech and music generators reached over the network with an API key. No download, no lane, no VRAM. They spend money at the vendor and the bytes leave the box, so a job under a local-only floor never reaches one. Add keys under SetupProviders. The rows show Docs and Pricing, not Download.

14.1Import a conversation export

1. Export from the other tool.

   In ChatGPT: Settings → Data controls → Export data. In Claude: Settings → Privacy → Export data. You receive a zip by email. Claude Code logs also import.

2. Open Bring your work.

   On the Archive, press ⇪ Bring your work, or drop the file anywhere on that screen. Also under SetupDataImport AI history or files.

3. Inspect the preview.

   Nightshift lists the recognised sources. A ChatGPT export keeps its full branch tree. Unknown content is marked opaque with a warning. Choose what to import.

4. Commit.

   Press Import. The commit refuses if the plan changed since the preview. Imports stream, so a 130 MB export commits without truncation.

5. Use it.

   Press Inspect sources › or Read it. Each exchange appears on the field as a historical attempt and in the Library. Ask for a working brief, or Combine with… other work and continue from it.

14.2Back up, restore and clear

Open SetupData. Backup, restore and clear run only when nothing is queued or running.

14.3Folders the models may read

The default workspace is Documents/Nightshift. Add more folders from any drive under SetupPreferencesFolders and pick which is default. Tools and watchers can never reach beyond the folders you grant.

15.1Routing presets

Under SetupPreferences, four presets sit above every slider:

Below them: latency target, quality target, cost ceiling, race width, premium budgets, cache aggressiveness, cross-check, answer fusion, answer judging, best-of-n on hard prompts, reasoning scaffold, adaptive compute, self-refine, scribe wave size, judge deadline, battery and memory thresholds.

16.1Mining and the proposals inbox

1. Mine a night.

   On the Jobs board inspector, choose mine it for ideas. A finished night or an import yields a candidate Method and Skill. Keep or dismiss each candidate.

2. Read the inbox.

   Open WorkersMining. The Workers tab shows a ⛏ badge when proposals wait. A strip above the Queue composer shows them too.

3. Accept or dismiss.

   Kinds: a job draft, a schedule or watcher, a Method or Skill, a fleet swap, a heal. Accepting a job fills the composer; nothing is submitted. A fleet swap waits for current work to drain. Dismissed proposals stay quiet for a month.

4. Turn on idle mining (optional).

   A background mining pass ships disabled. It runs only when the box is idle and yields the moment traffic arrives.

16.2Scribes and Analytics

17.1Three ways to run models

17.2Things to know

• When the allowance runs out, Nightshift offers a labelled replay of a real recorded night, or your own key.
• A background tab is throttled. The field shows tab_backgrounded. Keep the tab in front.
• Demo runs and private windows are memory only. Clear site data and your runs are gone. Export a night you want to keep.
• Mobile browsers are not tested.
• The footer may still print "RUNTIME :8088". There is no runtime port in the Web edition.

18.1How many models fit

18.2Which macOS

19.1Mac-specific steps

1. Install the verification pack.

   In Setup press Install verification pack. About 490 MB plus 296 MB of models, about 25 seconds. It installs the OCR reader and the contradiction guard. No terminal needed.

2. Put models on an external SSD (optional).

   Press Choose models folder… on the runtime card and pick the SSD.

3. Stop the runtime before you uninstall.

   SetupRuntimeStop runtime, then drag the app to the Trash.

20.1Operate the service

systemctl status meterlessd          # is the daemon up
sudo systemctl restart meterlessd    # restart it
journalctl -u meterlessd -n 200      # recent service log
tail -f /var/log/meterless/engine-0.log   # the engine host
tail -f /var/log/meterless/lane-<id>.log # one lane

21.1Windows-specific steps

1. Pin a GPU backend.

   Open SetupGPU. Read the decision. Choose CUDA, Vulkan or CPU. Press Restart runtime.

2. Install the OCR and contradiction pack.

   Run specialistd\install-windows.ps1 from the install folder in PowerShell.

3. Move the models folder.

   In SetupRuntime, choose another drive. The default is %LOCALAPPDATA%\ai.meterless.runtime\models.

4. Stop the runtime.

   SetupRuntimeStop, or run the command below.

   meterlessd --stop

22.1Reset things safely