Your standing work, done every morning by a crew that asks before it acts.

Free and open source. Runs on NVIDIA Nemotron 3 through Nebius Token Factory.

Open the appWatch the video · 2:40
$0.0216
one morning, the first run
$0.0209
the next morning, 3 % cheaper
24 %
saved by the router, 27 % the next morning
58,950
tokens across the crew in run one
23
model calls in run one, with 21 tool calls
3
sends you approved, each with a card
1
rule you saved, used the next morning with no card
16
memories after two mornings
0
replies sent without your OK

Measured on the recorded take of Oct 29: two real runs of the Morning brief, the second one starting from what the first one learned. Every number is in the release notes.

The plan, before anything runs · who does what and the price first
One row per agent · they work at the same time
Every step is open · click one, read what it read
Your morning, already read · inbox, calendar, GitHub on one page
Nothing leaves without you · approve, reject with a note, or allow from now on
One rule saved · you can revoke it anytime
What it learned · its own how-to, rewritten after each run
Run two, cheaper · it starts with what it learned
The 60-second tour · blur, focus, one thing at a time

Every clip is a recorded run on the real models.

Three doors to the same crew

One crew, one timeline, one set of rules. Start a run, approve it and read it from whichever door is closest.

Web/

The ops roomevery agent, every step and every cent on one timeline

Proof: the Checkout run run_668c7252 on production. Its three run_code calls each ran in a Nebius Sandbox VM (0.63, 0.75 and 0.71 s inside the VM), and the timeline marks each one with the Nebius Sandbox chip.

PhoneTelegram

Approvals as buttonsApprove, Always allow or Reject from your phone, and the web timeline records it

Proof: Mock-tested 40/40 against a stub Bot API (pnpm check:telegram). Real run from a phone: pending. Setup

Terminal and coding agentnimbus · nimbus mcp

The crew from your shellrun, approve and read it in the terminal, or let Claude Code or Cursor do it over MCP

$ NIMBUS_URL=…railway.app pnpm nimbus doctor

  • ✓Nimbusproduction answers · real Nemotron models
  • ✓Token Factory modelsthe key works · Ultra, Super and Nano listed
  • ✓Sandboxes whoamikey and project accepted · instance_max_timeout 3600
  • ✓Sandbox imagespython:3.12-slim and node:20-slim ready
  • ·TelegramNIMBUS_TELEGRAM_BOT_TOKEN not set (optional)
  • ✓Dataon the server · 7 workstreams, 7 with a SKILL.md
  • ✓Day captoday $0.31 of $1.00 · run cap $0.25

Proof: the rows above are nimbus doctor against production on Oct 29. The Release watch run run_e8b4dff2 was driven on production only through nimbus mcp tools by an MCP client script: five crew questions answered and one approval, $0.061. Transcript · Commands

What it does

  • Asks firstEvery side effect is a card. An approved one can become a rule, and you can revoke it.
  • LearnsMemories and a rewritten skill file after every run, so the next one asks less.
  • Priced per stepThree Nemotron sizes, routed by job, with the bill on screen while it runs.
  • Runs code in Nebius SandboxesThe code the crew writes runs in a disposable VM at Nebius, networking off, pandas and numpy ready: 3.5 to 4.9 s a call, under 0.8 s of it inside the VM.
  • YoursYour database, your keys, the tools you list. Nothing runs that you did not allow.
  • Always onSchedules in the app, your own Telegram bot, a CLI with MCP, and an OpenAI-compatible endpoint for OpenClaw and Hermes.
  • Open sourceMIT. pnpm dev runs it with no keys at all, on a scripted mock model.

How it uses Nebius

  • nvidia/Nemotron-3-Ultra-550b-a55bplans the run, routes each hop and reviews the result (the Orchestrator).
  • nvidia/nemotron-3-super-120b-a12bdoes the work: the Researcher, the Analyst and the Writer.
  • nvidia/NVIDIA-Nemotron-3-Nano-30B-A3Bchecks the output and writes the memories (the Checker and the Learner).

Every model call goes to api.tokenfactory.nebius.com, the OpenAI-compatible endpoint of Nebius Token Factory. The router picks the tier per hop, and the saving against running every hop on its ceiling tier is on the run summary.

Every run_code call goes to api.tokenfactory.nebius.com/sandboxes/v1, Nebius Token Factory Sandboxes: one disposable instance per call, from an image with pandas 3.0 and numpy, networking off, with the operation id on the timeline. The same key works for the models and the sandboxes.