Capability layer: a captain over a fleet of AI agents that builds and ships projects, with verification and approval built in — improve it by typing into it
Builders need a trustworthy command center for AI agents to manage multiple projects simultaneously
Loki provides a multi-user SaaS platform for commanding AI agent fleets across projects, with a focus on control, simplicity, and security
Direct the creation of everything you can imagine. Loki is the control plane for the age of autonomous creation: today one person commands a fleet of AI agents building software, tomorrow the same person commands fleets of robots building the physical world. Raw capability is no longer the constraint — human direction is. Every surface exists to raise the number of agents one operator can hold in their head and aim.
Leverage shifts from companies to individuals: a single builder, on their own machines and their own models, out-executes a team, and the only remaining bottleneck is their judgment about what to build. Deliberately open- and local-model capable, so creation is never gated behind closed frontier subscriptions. The commercial form — a high-retention SaaS for builders and small teams, with OrangeCat as the economic layer — follows from that goal rather than defining it.
Next.js 16 App Router · TypeScript strict · Tailwind 4 + ui-* design system (globals.css SSOT) · Drizzle ORM + PostgreSQL 17 (self-hosted Hetzner) · box-runner + Fleet Runner desktop (Electron, node-pty) · pgvector fleet-knowledge RAG (fastembed 384-dim) · Auth.js (GitHub/Google/X/email + OrangeCat OIDC)
Investigate this project with Loki →A Packages page. Under Fleet, Packages lists every shared package the fleet registers (bitbaum/fleet registers/packages.json) with its latest version and, for each, how many of your projects use it and how many of those are behind. Search finds a package even with a letter wrong. One page per package. It says what the package is, how to install it, where its code lives, and lists your projects that use it with the version each one is on — the ones behind first, saying by how much ("2 minor versions behind"), each linking to its project. "Built from" on every project. The Story tab now shows the shared packages the project's repository depends on, at the version its lockfile actually installed, with the ones behind marked and linked to the package. Nothing to set up in the projects. Loki reads each repository's package.json and lockfile from GitHub itself. It reads again when a page is opened and the last read is over six hours old, or when you press Refresh, and skips a repository whose code has not changed. A repository it cannot read is named, with the reason, instead of looking like it uses nothing. This is the read-only first step: choosing versions and rolling updates out come later.
Each project says how updates reach it. Under "Built from", one setting with three choices: Automatic (the default — an update merges itself once the project's tests pass), Ask me (each update waits for you on Approvals) and Stay on these versions (nothing updates until you change it back). The line under it says what will happen, and a change reaches the updates already open straight away, or within 15 minutes. Breaking updates always ask. A new major version, or a new minor under 1.0, waits for you whatever the setting, because the project's code may need changes to take it. Updates waiting for you are on Approvals. Each card says which project, which packages from which version to which, whether it is breaking, whether its checks pass, and links to the pull request. Update takes it (a breaking one is merged as soon as its checks pass), Not now hides it for a week, Stay on this version tells Dependabot to stop offering it. You can see every update in flight. "Built from" and each package's page show an open update under the package it changes: waiting on tests, held for you, or ready. Nothing new runs in the projects. The update pull requests are Dependabot's and the merging is the fleet's auto-merge, as before; Loki only puts a hold label on what should wait and takes off only the holds it put on — a hold someone added by hand on GitHub stays.
Before the first build, Loki says what the site will be built from. Under the start button, in plain words — "Your site will use sign-in, search, payments and email", how many other sites already run those parts, and that they are kept up to date automatically. It is read from what you wrote about the site; if it looks wrong, change the story — the build starts either way. The first build wires those parts instead of writing them again. Its instructions name each shared package, how to install it, what to use it for and not to write that feature itself — so a bakery's search box, pick-up times and Twint payments come from the same packages every other site uses. Client sites wear the client's look. For work done for a client, the build keeps the starter's design structure but replaces our colours and fonts with the client's, and does not install our design package. What was planned stays visible. Under "Built from", a "Planned" line names the parts the build was planned with that are not in the code yet. Whether the plan works is measured. Each first build records what it was planned with and, once its code has landed, what it actually installed.
No download for an account that has the cloud builder. Opened on a second computer, Loki said Fleet Runner was not connected “on this computer” and offered one fix: download the desktop app. “This computer” is the one machine where your account’s desktop app runs, so from any other device that offer was a trap — and the account’s cloud builder was online the whole time. Now the card under a dispatch says “Run it in the cloud” (one tap, the work moves), the empty terminal says “Use the cloud builder”, and the Control banner says plainly that there is nothing to install. The download is offered only to accounts whose cloud builder is private. The Terminal no longer says "start it" to a builder that is down. A deep link to a project with no session read “isn’t running on Cloud builder — start it, and it opens here” while the chip beside it said the cloud builder was offline. Now that screen says the builder is offline and what happens to the work, and the start offer waits for a builder that can take it. The cloud builder's pulse is public. /api/health/builder answers 200 while the shared cloud builder is connected and 503 when it is not, so the box watchdog and the fleet sweep alert on it the way they already alert on a dead site. It was offline all night on 2026-10-11 and only a chip inside a signed-in page knew. Each deploy now prints the builder's state and its last journal lines into the CI log, and the feedback queue says “Cloud builder offline” instead of “Watch shows why” when that is the case. The Mac download is not damaged. macOS says the app “is damaged and can’t be opened” because the build is not yet signed by Apple, and right-click → Open does not clear it. /download now leads with the .zip and one Terminal line that installs, clears the refusal and launches it. Signed builds need an Apple Developer account (USD 99 a year); the pipeline is ready for the certificate. A Suggestions page. The "Build next" row on the Loki start screen only knew about code. The new page under Now reads every front Loki keeps records on — Build, Business, Goals, People, Money, Habits — and offers one thing to do per fact: the report nobody fixed, the delivery waiting on your yes, the goal due Friday at 40%, the subscription renewing Tuesday, the friend you have not spoken to in forty days, the twelve-day streak not checked off yet. Every card says which record it came from. Two answers, no typing. "Do it" hands the ask to Loki, already written and already scoped to the project, or opens the one place only you can do it (checking off a habit, accepting a delivery). "Not now" puts it away for a month, with an undo. What you put away stays away on the start screen and on Control too. It keeps score, from records. Taken this week, days in a row, taken in all — counted from the decisions you actually made, never estimated. Private fronts stay private. People, Money and Habits are read only while your PIN zone is unlocked; locked, the page says so and offers the unlock.
Work sent to the cloud starts again. Since 22:25 UTC on 2026-10-11 the cloud builder stopped itself a few seconds after every start — 220 times — so every cloud dispatch waited in the queue. The form-filling window added to Fleet Runner that evening belongs to the desktop app, and the cloud builder, which runs the same code without a screen, tried to load it at start-up. It now loads only when a form arrives, which only happens on a computer. A new check fails the build if anything the cloud builder loads ever needs the desktop app again. Fleet Runner 0.8.42.
Search forgives a slip. ⌘K, People, Projects, Control's "Find project", Prompts, Goals, Events, chats, essays, the model store and the project pickers all matched the exact letters typed, so "lkoi" found nothing and "jonh" found nobody. Now one wrong, missing, doubled or swapped letter still finds the thing (two in long words), accents need not be typed ("zurich" finds Zürich), and "orange cat" finds OrangeCat. The best match comes first in ⌘K, People, Prompts and the pickers, and ⌘K and the prompt picker mark the part of each result that matched. Lists with an order of their own — the sort you picked on Projects, Control and the model store, deadlines, recent chats, the goal tree — keep it; the search only narrows them.
Every agent Loki sends into a bitbaum repo is now told what already exists. Alongside the project's own profile, the dispatch carries the studio's rules: read fleet's AGENTS.md, STACK.md and SHARED.md first, start a new product from the fleet app template, use the shared packages (each with its one-line purpose and install command) instead of writing another health route, model call or set of security headers, keep the security bar, and count work as done only once it is merged, deployed and checked live. The package list and the stack are read from bitbaum/fleet when the agent is sent, so they are never a copy kept in Loki. If GitHub is slow or down, the dispatch goes ahead with the last saved list. Repositories outside the bitbaum org get nothing extra.
New sites start with accounts, three languages and strict security. Both ways Loki creates a site — "Make it happen" with the Next.js starter, and a new site on the box — now produce the same app: the bitbaum starter. People can create an account with a password, a sign-in link or a passkey (and with OrangeCat once the site is registered there), add two-factor sign-in, and use it in English, German or French. It ships with strict security headers, a health check, tests and the pipeline that deploys each change. A new site gets its own database. Its accounts live in a database made for it on the server, with its secrets generated there and never shown.
"Build next" on the Loki start screen. Under "What are we working on?" there is now a row of things you could build, picked from your own records: visitor reports nobody has fixed, a project with code but no live site, the next step your last session left, a goal stuck at zero, an idea from your sticky notes, and needs people have posted on OrangeCat (the ones closest to what you already build come first). Each one names why it is there. One tap starts the work. Tapping a suggestion for one of your projects sends it to an agent on that project. A new idea opens the new-project form already filled in, which leads straight into a kickoff. "Other ideas" shows the next set. With nothing set up yet, you get first builds that each end in something live: your own site, a page for your offer, a booking page. Loki pushes to get it built. Ask "what should I build?" and Loki answers from the same list, recommends one thing and why, scopes the smallest first version, and ends with the exact line that starts it.
Visual fixes get a verdict too. A layout, a colour, a thing that overlapped: page text could not say whether the fix worked, so those were the ones still sent to your eyes. Now, when text cannot tell, Loki takes a phone-sized picture of the live page — scrolled to what the reporter pointed at — and asks a vision model the same question. The card says which it did: "Loki read /pricing" or "Loki looked at a picture of /pricing". You see what Loki saw. When even the picture cannot settle it, the picture is on the card, under the reason, so "It worked / Not fixed" can usually be answered right there. Tap it for the whole page; it is also in every row's details. A server with no browser says so on the card, in words, instead of a silent "cannot tell".
The evening plan now actually arrives. The 19:00 proposal for the night shipped yesterday, but its timer only reached the server when someone ran a setup command there by hand, so no plan was made. Every deploy now installs the scheduled jobs itself, and a test fails if that ever stops. Fix checks have their browser. The picture Loki takes of a live page needs a browser on the server. The deploy installs it too, so the first visual fix gets a real verdict instead of "no browser".
A finished change comes with proof. Agents now report shipped work as the live link first, screenshots of the real deployed page (phone first), and a plain list of what was and was not checked. pnpm run ship:report produces it: it confirms the site is running the shipped build and photographs the page, instead of announcing a merge and hoping. Loki remembers how you use "Build next". Each time the row is shown, tapped, or rotated past is recorded, and Loki reads it back: ask "how am I using Build next?" and it says what gets tapped and what it keeps offering for nothing — and it uses the same record when you ask what to build, so it leads with the kind of suggestion you actually take.
Unfunded keys save. A key the vendor recognises but cannot bill yet — no credits, a spending limit reached — used to be refused as a bad key. It is now saved with the vendor's own words and the link to its billing page, and Loki uses it the moment the vendor does. The first real xAI key pasted into Settings hit exactly this wall. What your keys spend. Each key's row in Settings → AI shows today's and the last thirty days' tokens and calls, and links to the vendor's credits and spending cap. Tokens, not francs: the vendor's meter is the bill. Scale, written down. docs/architecture/models-and-money.md says why the free pool is a tasting menu, why your own key is the path that scales to a billion builders, and what a pass pays for.
Autopilot overnight, with a budget. While you sleep, Loki builds from your feedback — one fix per project, your own notes first — reads one site that has not been read this week and files what it finds, and files away reports nobody started in 21 days with the reason on the row. The limit is a number of runs a night, in Settings → When an agent finishes (default 4, 0 keeps the filing and stops the building). A project paused on Control is left alone. The idle nudge that woke three projects a night on a generic "next best" guess is retired: a run with no brief is where the queue of things nobody asked for came from. The morning note. /feedback opens with one line on what the night did and what it spent where the runner reported a price; the same line arrives on Telegram and push. Nothing when the night did nothing. Feedback on a phone. The question — how many need you — leads the page; the pickers are one quiet line; the numbers close it. Loki reads the page. The widget's Watch measured — small type, tiny targets, missing names — and said so on every page while the page in front of you said nothing first. It now also reads each page once per visit as a first-time visitor on a phone (purpose, hierarchy, content, fit, function) and says at most three things, each quoting the page, or nothing. Same "Fix this". Essay: /thoughts/loki-reads-the-page. Which model are you. Loki's chat says what it is running on from its own prompt, in one sentence, instead of a tool round and a referral to Settings. Share on OrangeCat. Every Thoughts essay's share bar leads with OrangeCat: one tap opens the composer there, prefilled, on the reader's own account. "Back in 30 minutes." Say it to Loki and the next screen you open — Control, or the chat — leads with what happened while you were away: what needs you, what went live, what finished and what still runs, each line a link to where it is acted on. One tap and it is gone. Nothing is recorded during the absence; the summary is read from what moved. Fewer things thrown together. A feedback row's context is one sentence that wraps like one; the pickers are words, not boxes; "Pointed at" reads as a sentence; the thumbs-down form in the chat is part of the thread, not a second dialog; a review files the screenshot it judged with its first finding, so you read the finding beside the picture.
A model call is bounded by silence, not by a stopwatch. A 30-second cap on the whole call killed the first real turn on a user's own key: grok-4.7 thought for longer than that before its first word. Now a vendor may take up to 90 seconds to say anything on your own key, may stream for up to five minutes, and is only cut off when it goes silent for 30 seconds — and the sentence you get says which of those happened. Your model failed? One tap answers on the free models. Under such a turn a button sends the same words on Loki's free chain, for that turn only. Loki never switches silently: your key failing is the one thing you need to see. Every key action says what it did. Adding a key reads "xAI · grok-4.7 added. Loki starts here now. Testing it…" and then "Answered in 1.3 s — ‘Hello, I'm Grok.’": a real, tiny turn on the key, timed, so "added" is backed by an answer. A Test button on each row repeats it; changing, re-ordering and removing say what changed. Screens that lay out on a phone. The model picker's rows were being squeezed in the bottom sheet and overlapped; a key's row in Settings wrapped its model name letter by letter. Both fixed; "xAI" is no longer shown as "XAI". No more "update password?" The key field is masked but is not a password field, so the browser stops offering to save your API key as the site's login. Your words appear as you speak. Press the mic and the words scroll in as they are heard, one line, dimmed while the server transcribes and replaced by its words when they land. The checkmark still ends the take. Where the phone's recogniser cannot run, the wave stays and the recording is untouched. (chatkit 0.7.0.)
Models, compared. /models lists every provider you can bring — OpenRouter, Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral, Groq, Together, Cerebras — with who they are, what they are known for, what they give away, and their newest models with live prices per million tokens and context, read from OpenRouter's public catalogue and refreshed on a button. A table sorts them by what decides a bill. Add to Loki opens Settings with the provider chosen; Use chooses the model too. The free pool, in plain words. The models Loki uses at no charge are shared by everyone here and rationed per person; when a day's allowance is spent it is spent for everyone. The capacity counters now sit behind one line instead of twelve cards. Your own key does not draw on the pool. A second provider. With one key connected, Settings says why a second matters: Auto gets a choice when the first is slow or out of credit, and you can compare two models on the same question.
Nothing runs at night that you have not seen. Every evening Loki shows tonight's plan with names and a price on it — which fixes, which site, how many of your runs, and what it would cost going by your own last runs — and asks for one tap: Approve or Reject, on Control, on Telegram and under Approvals. A plan left undecided expires at 02:30 and the morning note says nothing was built. Before this, the night decided for you; the plan was only visible the morning after. Allow it in advance, for as long as you say. Settings → Autopilot: "Nights run without asking" for a week, a month or three months, optionally under a dollar cap a night. Under an allowance the evening message says "runs without asking until …"; a night over the cap asks again. Free things stay free. Filing away reports nobody started in three weeks and moving stuck rows to the cloud spend nothing, undo in one tap, and happen with or without a yes. The evening plan and the morning note are assembled from counts — no model is called to write them. The late-night "go to sleep, autopilot has it" strip on Control is replaced by the Tonight card, which shows the plan instead of offering a switch.
The owner's launcher is one thing, always in one place. On your own site it was two: "Loki · watching" in a corner with a ◎ beside it, and a "Loki is watching" bar at the top that took over whenever the corner was busy — on a phone, every scroll. Now the launcher never hides and never fades for the owner, the ◎ is its other half (there whether watching or paused), and the bar appears only on a page that hides the launcher outright — carrying the same ◎. Show Loki this screen. ◎ now offers This screen beside tapping an element: it marks what is on screen right now, the way a screenshot would frame it, and the conversation opens with it attached — say what you would change, ask Loki, or build it.
The whole landscape on /models. Every chat model in OpenRouter's public catalogue — about 460 from 60 labs, Kimi, GLM, Qwen, Llama and MiniMax among them — with open weights marked, an intelligence index, the price in and out, context, when it was listed, and every way to run it through Loki. Filter by open weights, free to try, sees images, reasoning, lab or a price ceiling; sort by best value, smartest, cheapest, newest or longest context; the filters are the URL, so a view is a link. Use switches a model on a key you hold; Add key opens Settings with the vendor and model chosen. "Check for new models" names what appeared since the last read. The Chinese labs directly. Moonshot (Kimi), Z.ai (GLM), Alibaba (Qwen) and MiniMax can be connected with their own keys, not only through a router — and Fireworks as a host for open models. A model on your own machine. Point Loki at Ollama or LM Studio behind a tunnel (Tailscale Funnel, cloudflared, ngrok) and it thinks for free, with your data staying on your laptop. The recipe is on the store; Loki checks the address, lists the models it finds, and only ever connects over https to a public address.
A greeting never spends a frontier model. With your own keys, every turn is judged light, standard or heavy before a model is chosen, and Auto picks the cheapest model that is good enough: a cheap one for lookups and captures, the best value for ordinary asks, the smartest you have for code, strategy and anything you ask to be thought about carefully. The footer says which turn it was and why. Your stance. Thrifty, Balanced or Best — how far up Auto may reach. Even Best sends the small things to the cheap pick. Your picks. On /models and in Settings → AI you see the three models Auto chose and why, change any of them to another your keys reach, and hand it back with "Loki decides". The composer's Auto row names the two ends: which model takes the light turns and which the heavy ones.