Bring your context
Point Ultron at exactly what matters: your GitHub repository, the files you upload and the tools you already use.
Sessions
Equipped to do
real engineering
Ultron edits files, runs terminal commands, searches the web and more, in a sandboxed machine of its own for every conversation.
Open the appOpen the appHand off the work, keep the decisions
A session takes your files and a skill from the Command center, does the work in its own container, and comes back with something you can use.
Built around the way you work, from a quick question to a long job you walk away from
Start from a skill
Type a slash for the Command center and its seventy-nine skills, or say what you need in plain language.
Choose the effort
Minuet for speed, Allegro for almost everything day to day, Plinth for long jobs you walk away from, Forte when being wrong is expensive.
Watch every step
Every command, search and edit shows in the conversation as it runs, with what it found and what it changed.
Keep it running
Hand off a long job and it keeps running in the background, while the Control Center follows every running and recent job.
Pick up any session
Pin the ones you come back to. Projects keep conversations, files and memories together, and any conversation branches from any message.
Every result opens right beside the conversation
Pictures open in the media preview, videos in a full editor, notes in their own editor, apps on Ultron's computer and every change in the Code panel, all beside the chat.
Media preview
Open a picture and step through the whole set at full size. Zoom, rotate, crop or black out a detail, then copy or download it.
Notes
A full editor for your notes: headings, lists, checklists, quotes and code, every block a slash away.
Code
Every conversation keeps a repository of its own. Each commit opens on its diff, and with GitHub connected it lands on your branch.
Video editor
Every video in the conversation opens in a full editor. Trim it, split it, frame it for the feed and attach the cut to your next message.
The computer
Ultron's own browser runs the apps it builds and the pages it works on, and you can watch every step.
Live apps
Every app Ultron builds goes live at its own address, and an update redeploys it in place.
The harness
Ultron finishes the same task on a fraction of the model calls, by never spending one on work a script a workflow or the cache can finish correctly.
Not by picking a smaller model. Data shaping runs as a script on the pod, multi-step jobs run as workflows and repeated work comes from the cache, so the model is left with only the judgment calls that need it.
- Script on the pod
- Workflow
- Cache
- The model
How the cost comes down
Three things bring it down, and none of them is a better model.
Fewer calls
The harness plans the whole task before it starts, batches what it can, and stops asking the model to do anything a function can do on its own.
Offload the rest
Anything deterministic leaves the model entirely. Data shaping runs as a script, multi-step jobs run as workflows, and repeated work is served from cache.
Compile what repeats
Work that recurs runs as a routine verified once and invoked by name, and its result is checked against its invariants before anything downstream trusts it.
Work per dollar, not price per call
The unit that matters is not what a call costs. It is how much of a task one call finishes.
The roundtrip tax
Most steps in a long task are searches, transforms, lookups and formatting between the few moments that need thought. A full roundtrip for each is the tax the harness removes.
The ceiling is honest
Reasoning cannot be compiled away. On repeated work nearly every step is deterministic, on original design work almost none is, and both are true at once.
Caching is not enough
A cache cuts the price of a repeat. It does nothing about the count of first-time calls inside one task, which is where the bill comes from.
The route
Where the work runs. Almost none of it ends at a paid model.
Chat, automations and canvas all funnel into the same four surfaces. A step that a script can do correctly is a step the model is never asked about.
ChatAutomationsCanvas
Planned onceRouted step by step
Script on the podfreeWorkflowfreeCachefreeThe modelbilled
Compiled, then checked
The tenth time a task shape appears, nothing about it is new. It runs as a routine verified once, and the result stands only if it survives a check.
Recognise the shape
Recurring work is matched to a routine that already exists and is already verified, so what is left to decide is a handful of parameters, not a whole procedure.
Run it as code
The routine runs on the warm pod, where every tool is a local command. A loop over a list of records is one execution, not a roundtrip per record.
Escalate on surprise
Output is checked against what the routine promised. Anything that fails, or looks unlike what was expected, goes back for real reasoning instead of being passed along.
Where the cost leaks
Five patterns account for most of what a task overspends, and every one shows in the trajectory log before it shows on an invoice.
Re-sent history
Prune before the call and hold the stable prefix in cache, so a step pays for what changed rather than for everything said so far.
Raw payloads
Parse, filter and aggregate on the pod. Only the extracted fields travel back into the conversation.
Repeat work
Results are cached by content, so a second read of the same file or endpoint resolves without a call at all.
A roundtrip per step
Deterministic runs of steps compile into one execution and come back as a single structured result.
Blind retries
Check the invariants, escalate once with what went wrong, and never re-run the same failing shape hoping for a different answer.
Cost per successful outcome
Spend per call is the wrong denominator. What matters is what a finished, correct outcome costs, counting everything spent on the attempts that did not finish.
Count every attempt
A task that takes three attempts to stick costs three times the tokens and the tool calls, and whatever it takes to undo the first two. Spend per attempt hides all of that.
It changes the decisions
Measured per outcome, reliability stops being a separate virtue from cost and becomes the same thing: the fastest way to make a task cheaper is usually to stop it failing.
It prices the invisible half
Tokens burned on a dead end, tools billed for work that was thrown away and the minutes spent cleaning up all belong to the outcome that finally shipped.
What gets measured
A number nobody can check is a claim, not a result. Every run is measured on three layers, and nothing counts as a success unless the third says so.
Accounting
Input against output, what the tools themselves billed, and how much of each prompt was served from cache rather than paid for again.
Trajectory
How many steps it took to finish, how often the same file was read or the same endpoint queried twice, and what was spent keeping the conversation inside its window.
Verification
A hard pass or fail decided by code: assertions that run, a build that completes or does not, the state of the record afterwards. Never the run's own account of how it did.
Why the others cannot follow
Claude Code, Codex and Gemini CLI are good harnesses, shipped free by the companies that sell the tokens they spend. That decides which optimisations each of us can even consider.
Sold by the token
When the tokens are the product, every step moved off the model is revenue removed. A call made cheaper sells more calls; a call that never happens is lost revenue, so it stays at the bottom of the list.
Sold by the month
Ultron charges a subscription, so inference is a cost line rather than the product. Every step that runs as a script, a workflow or a cache hit is margin kept rather than revenue forfeited.
The layer above runs on subsidy
Tools that resell inference at a flat monthly price can find their best customers are their most expensive ones. Spending less inference per finished task is the only version of this that survives its own success.
The point is not that the model is cheap. It is that the model does less.
A step that runs as a script is a call that never happens. A step served from cache is a context that never gets re-sent. Judgment is the floor, and everything above it was roundtrips, re-sent history and retries, which get fixed by building.
Frequently asked questions
What is a session?
A conversation with a machine of its own: a sealed micro-VM, not a shared container, where Ultron edits files, runs terminal commands, searches the web and builds.
Does a session keep working while I do something else?
Yes. Long jobs run in the background, and the Control Center follows every running and recent one. When a job finishes, Ultron picks the conversation back up with the result.
Where do the files and apps end up?
Every file a run delivers lands in your Ledger as a set named after the run, and every app goes live at its own address. With GitHub connected, each change also lands on your branch.
Which effort should I pick?
Allegro is the sensible default for almost everything you do day to day. Minuet answers almost the moment you ask, Plinth is the one to hand a job to and walk away from, and Forte is for when being wrong is expensive.
Can I go back to an older session?
Every conversation is kept. Pin the ones you come back to, keep related ones together in a project, and branch any conversation from any message.
What is a skill?
A unit of work Ultron can dispatch, scoped to the tools and the model tier it needs. Trigger any skill from chat with a slash command or plain language.
Can I build my own skills?
Yes. Scaffold a skill with a manifest, a tool allow list and a model tier, and it loads on the next start.
Which models does Ultron use?
Claude, GPT, DeepSeek, Kimi, Grok, GLM and Qwen, wired into one runtime. Pin one per skill or let the router decide on every turn.
Does Ultron act without asking?
No. It drafts a plan, runs the skills and pauses at every decision that leaves the building. Nothing outbound happens by default.
What does Ultron remember?
Anything worth keeping lives in the Ledger: the notes you write, the memories the agent saved, your files and signed agreements. The next turn retrieves the right items on its own.