A 3D office where a hierarchy of LLM agents actually does the work.
One application, thirteen employees, and a cost ceiling on every run.
You describe a task to a CEO agent. It builds an objective, puts specialists on it, has them argue the approach, gets a file-level plan, writes real code, and tries to break it — every step a separate model call, priced for the work it actually is.
This is the application, not a diagram of one
Every image below is a capture of a live instance — the same build the quick start produces, running with thirteen employees and a real brief in flight. Pick a surface to see what it does.
The office is the application. Employees sit at named desks and change colour as their status changes, the panel on the left is the live state of the organisation, and the dock is where work is commissioned.
Twelve surfaces of the console, captured at 1600×1000 from 127.0.0.1:8787.
Five stages, from a sentence to a reviewed change
Complexity, prompt, turn, stage, run. Each stage is a group of LLM calls with its own mode, its own roles and its own model budget — which is what makes a thirteen-person company affordable to run.
- 01
Intake
CEOA nano model reads the brief, sizes it, and decides how much company the task deserves. This is the cheapest call in the run and it shapes every one after it.
- 02
Research
Specialists, in parallelTwo to four specialists read, argue and record what they found. A debate stage has them take positions and a chair summarise what survived.
- 03
Plan
CTOA frontier model turns the surviving position into a file-level plan: which files change, in what order, and what would falsify the approach.
- 04
Build
DevelopersDeveloper agents write real files into the run’s workspace, through tools whose path arguments cannot leave its root.
- 05
Break it
QA, then the CEOQA tries to falsify the result. The CEO reports back with what was done, what was skipped and what is still unproven.
The record is the run, not a summary of it
Every turn is kept: which model answered, what it cost, what it was asked and what it wrote. When a run goes wrong you read the turn that went wrong instead of guessing at it.
- Every turn is a separate model call, priced and attributed — so the cost of a run is explicable rather than one opaque number.
- Approvals stop the office until a human answers, and one that times out counts as refused rather than as permission.
- Cancellation is safe, and the transcript keeps what happened up to the moment you stopped it.
Built for the shape of the work, not the shape of a chat box
The office is a small number of working parts. Each one links to the detail, including what it cannot do.
Cost-aware routing
Every role declares a tier band. The router picks the cheapest model inside it that can actually do the job, so a summary never costs a frontier call.
Read moreFour stage modes
single, parallel, debate and review-loop. The shape of the work decides how the agents meet, so a two-line question and a migration do not get the same round table.
Read moreTools that stay put
Fifteen built-in tools, eight of which take paths and resolve them through one choke point bound to the run’s workspace root. run_shell and git are not confined and are approval-gated instead — and the page says so rather than implying a boundary that is not there.
Read moreA real plugin system
Add a provider, model, skill, routing rule, tool, role, pipeline or panel without rebuilding the application. Eight kinds of contribution, all live.
Read moreA floor that grows
Each organisation gets its own floor. Space is generated from a block kit and extends itself when it runs out of desks.
Read moreRuns without keys
With no provider configured the whole pipeline still runs against scripted employees, so the office is demonstrable before it is billable.
Read moreThe cheapest model that can actually do the job
A one-line intake summary and an architecture review are not the same purchase. dev3d routes them to different models, inside a band the role declares.
- Roles declare a minimum and maximum tier — a policy the router cannot exceed.
- Three postures: cheap, balanced, quality. Switchable per installation at runtime.
- Providers with no key are catalogued but excluded from the routing pool, so a missing key cannot silently fail every turn.
- Plugins can contribute routing rules scoped to a task class — a preference, never an override.
Extend the office without modifying it
A plugin adds models, skills, routing rules, tools, role templates, pipelines or panels as validated data. This site serves the catalog a dev3d instance installs from.
Office echo
dev3d.office-echo
A minimal code plugin: registers one tool employees can call, and logs run lifecycles. It exists to prove the code path and to be copied.
Local coder
dev3d.local-coder
Registers a whole provider: a local OpenAI-compatible model server, its endpoint and the model it serves. The keyless loopback base URL is what tells the off…
Cost guard
dev3d.cost-guard
Keeps the cheap work cheap: routes intake, summarising and research to the least expensive capable model, adds a small local model to the catalog for it to a…
Written down, including the parts that are awkward
Installation, concepts, guides and reference. The reference documents limits as readily as features — the plugin that cannot be unloaded from memory, the confinement boundary that is a boundary and not a jail.
Introduction
What dev3d is: a 3D office where a hierarchy of LLM agents plans, builds, reviews and reports.
/docs/introductionModel routing
How one model is chosen for one turn: capability filters, tiers, a weighted score.
/docs/concepts/model-routingBuilding a plugin
Declarative or code, what a manifest may contribute, and the rules the host enforces.
/docs/guides/building-a-pluginpnpm install # Node >= 24
cp .env.example .env # optional: add provider keys
pnpm dev:server # orchestrator on http://127.0.0.1:8787
pnpm dev:web # office UI on http://127.0.0.1:5273Run it yourself in about two minutes
dev3d is open source under the AGPL. Install it, run the mock pipeline, and decide whether to point it at a real provider once you have seen the shape of the work.