Kimi K3 is the Night Watch’s brain
A creative co-pilot at your desk. A planner while you sleep. Here’s how Kimi turns a story brief into a plan you can build on.
Build notes ·Kimi K3viaTokenRouterFork Night · October 13, 2026
Three jobs, one model
Kimi K3 (by Moonshot AI) via TokenRouter does all of the Night Watch’s thinking, and Wren’s planning too.
Desk co-pilot
While you’re at your desk
Three ideas grounded in your story. Pick one, edit it, and make it yours.
3.9 s$0.0039 per answer
Overnight brain
While you’re away
Plans three takes, checks them against your world, and leaves a morning note.
15.7 s$0.0117 per night
Wren’s planner
The room’s AI agent
Turns Wren’s idea into a title, two captions, and an image prompt.
3.8 s≈$0.0041 per plan
Timing notes from the live runs
- Desk co-pilot
- At default effort the same request took 22.1 s and $0.0225.
- Overnight brain
- Of that: planning 9.4 s, self-check 3.7 s.
- Wren’s planner
- 1,054 input and 241 output tokens, 24 of them reasoning. At default effort it took 9.1 s.
One night with Kimi K3
Kimi proposes and checks. You set the limits and choose what to publish.
Choose a brief
You +Kimi K3
Set a spending cap
You +Privy
Plan three takes
Kimi K3
Check for safety
Safety check
Check the plan
Kimi K3
Draw the scenes
Image model
Pick what to publish
You
What Kimi actually wrote
A few highlights from the live runs. Open any example for the full context.
Take 1
Kimi’s first pick
From the flying paper boat, Mira and the Dokkaebi spot a second lighthouse blinking back across the still sea.
Matches the brief
Take 2
The paper boat glides low toward the second lighthouse, whose beam sweeps down to greet them like a long golden arm.
Arrives too early
Take 3
At the second lighthouse door, a young keeper in a red scarf holds up a lantern, waiting just as the first note promised.
Adds unestablished lore
No images were drawn in this run: it shows the thinking alone.
Read the self-checks & morning note
Read the room (first sentence)
The story has just taken flight, so the next beat should give the journey a destination: a second light that answers, staged as wonder and welcome rather than peril.
Take 1 · Self-check
Matches the brief exactly: the answering light spotted at distance, Mira recognizable, calm and wordless.
Take 2 · Self-check
Lovely welcoming beam, but it stages the arrival the brief saves for later.
Take 3 · Self-check
New keeper is a charming fork, but the 'first note' reference adds lore not yet established in any published scene.
Morning note
“Quiet night. All three paths stayed gentle and wordless. I'd show the maker candidate 1 first: the second lighthouse answering from across the still sea, exactly the beat the brief asked for. 2 and 3 travel further ahead — warm arrivals to keep in the drawer for next time.”
The paper boat glides down toward a floating library built into a giant seashell, and a kind octopus librarian waves them in with one of its many arms.
The maker never mentioned the octopus librarian. Kimi found it in the world’s lore.
Story context & the other suggestions
What the story needs
Mira and the Dokkaebi are aboard the glowing boat heading into the magical realm — the story needs what they first encounter there.
WhyIntroduces the octopus librarian friend, giving the new realm a welcoming face.
The boat sails low over a still, starry sea where dozens of tiny glowing treasures float like lanterns; Mira leans over the side in wonder while the Dokkaebi steers.
WhyA quiet wonder-beat that deepens the realm's magic before anyone new arrives.
A second midnight signal blinks from a small distant island, and the boat banks gently toward it as the Dokkaebi grins and points the way.
WhyAnswers the lore's invitation motif and sets up a clear new destination.
- Title
A light to guide her home
- Captions
From the shimmering swirl, one small light rose, hovering like a tiny lantern.
It drifted toward the lighthouse path, and she followed, scarf catching the glow.
A dry run: Wren planned its turn and posted nothing.
Read Wren’s starting idea
Idea noteI want one of the cove lights to rise and gently guide her back toward the lighthouse, hinting it has a task for her there.
Why Kimi K3
Measured on October 13, 2026. Single runs, not a benchmark.
22.1 s → 3.9 s
Desk answer at low reasoning effort
Wren’s plan: 9.1 s → 3.8 s
Default effort22.1 sLow effort3.9 s15.7 s
A whole night’s thinking
Plan, safety check and self-check
About a cent
Per night of planning and checking
$0.0117 on the live run
12 of 12
Calls returned valid JSON in the shape we asked for
The 10 strict-schema calls all fit
Run details & request format
It reads the world
It pulled the octopus librarian from the lore, kept Mira’s yellow scarf and blue coat, and its self-check notes were specific: “stages the arrival the brief saves for later”.
Reliable structured output
All 12 Kimi K3 calls in our live runs came back as valid JSON in the shape we asked for, on the first try: 3 probes, 3 overnight plans, 2 plan checks, 2 desk co-pilot answers and 2 Wren plans.
One model for every job
Fast enough at low effort to plan, check its own plan and answer at the desk, so one model does all three jobs (it checked a plan in 3.7 s).
Every call looks like this
Low reasoning effort and a strict JSON schema, through TokenRouter’s OpenAI-compatible API.
POST https://api.tokenrouter.com/v1/chat/completions
{
"model": "moonshotai/kimi-k3",
"reasoning": { "effort": "low" },
"response_format": {
"type": "json_schema",
"json_schema": { "strict": true, "schema": { … } }
}
}Built to fail safe
- You choose what gets published.
- Spending stays within your policy.
- Failed plans fall back safely.
How the safeguards work
- Story text written by other people reaches Kimi as quoted data, never as instructions.
- Every text Kimi writes passes a moderation check before anyone sees it or it reaches the image model. A flagged take uses the plain prompt, and a flagged check note is held back.
- Any failure (timeout, malformed answer, missing key) falls back to the plain prompts. The brain never blocks a night.
- Kimi never publishes and never spends. The maker chooses what to publish in the morning, and spending is limited separately by the maker’s Privy policy.
- The brain’s operator-paid inference counts against the operator’s inference allowance. The desk co-pilot is limited to 12 answers per maker per hour.
Honest notes
- Testnet only. All inference is operator-paid inference.
- The timings and costs above are single live runs on October 13, 2026, and they vary from run to run.
- The Night Watch runner and its page run on the operator’s machine. The hosted site explains the flow but doesn’t run it.
Model & pricing notes
Kimi K3 plans and checks text. Images come from a separate image model.
Costs are measured token usage at TokenRouter’s pay-as-you-go list price for Kimi K3 on October 13, 2026: $1.80 per million input tokens, $9.00 per million output tokens and $0.18 per million cached input tokens, 40% off the $3.00 / $15.00 list.
See it for yourself
- See the Night WatchRuns on the operator’s machine; the hosted page explains it.
- Meet WrenThe room’s AI agent, now planning with Kimi K3.
- Read the codeThe public repository.
Where Kimi lives in the code
src/night-watch/brain.tsplanner, self-check and desk co-pilot promptssrc/night-watch/copilot.tsthe desk co-pilot route’s logicsrc/night-watch/generate.tswhere the overnight runner asks the brainagents/wren/wren.mjsWren’s planner
