Features
Everything happens in the office.
Hire agents, seat them, give them skills and rules, then watch the fleet work — every capability traced back to the chair each agent sits in.
The HQ floor: a dossier on the left, the cast walking the floor, the roster and activity feed on the right.
Build the org
Six panels. One chain of authority.
Roles, teams, skills, rules, servers and voices — each one attaches to a seat or an agent, and each one is gated by clearance.
01 · Roles
Define a job once. Every seat inherits it.
A role carries a title, a discipline and a default clearance. Seats reference it and inherit live, so editing a role moves every seat that holds it.
- 217 professional roles across 21 disciplines, seeded conservatively — nothing starts at 6+
- Disciplines are free text: type a new one and the category appears
- A role held by any seat cannot be deleted out from under it
02 · Teams & seats
The org chart is the permission model.
A team lives on a floor, has a mission and a set of seats. One seat is the leader — the only one that can delegate, and it is labelled so you can see it at a glance.
- Seats inherit clearance from their role, or override it explicitly
- Teams nest under a parent, forming a strict downward tree
- Team memory is shared by everyone on the team and survives seat changes
03 · Skills
Teach a seat how to do a task.
Skills are markdown playbooks injected into the run prompt. They belong to the seat, so whoever sits there picks them up.
- A minimum clearance keeps a skill off seats too junior for it
- Written once, assigned to as many seats as need it
04 · Standing rules
How we always work, written down.
Rules are prepended to every run of the seats in scope, ahead of their skills. A seat follows the global rules, its team’s rules and its own.
- Scoped global → team → position, with an explicit order
- Pause a rule without deleting it
05 · MCP servers
External tools, gated like built-in ones.
Register an MCP server once, then assign it to seats. Its tools merge into the run’s tool list and pass through the same executor gate as everything else.
- A minimum clearance per server
- Remote HTTP transport with an optional bearer token and extra headers
06 · Personalities
The one thing that belongs to the agent.
Everything else comes from the chair. A voice should not — move Alden to another desk and Alden still sounds like Alden.
- An identity stance: asked if it is an AI, it says so in one sentence and stays in character
- Re-sent on every retry, so a long run never drifts back to the model’s default voice
- All fifteen of the cast ship with a voice
Hire
Bring a new agent into the building.
An agent needs a dossier and two images: a portrait and a walk-cycle sprite sheet. It arrives at clearance 1 — what it can really do comes from the seat you give it.
- Sprite sheets are validated: PNG, exactly 192×192, a 4×4 grid of 48px cells
- Pick a floor and a free seat — two agents can never share a desk
- Per-agent overrides for reasoning, image and speech models, web mode and budget
Operate
The whole fleet, one screen.
Every agent as a tile with its live status. Open any agent’s terminal to watch its event stream scroll by, or step in to approve, stop or redirect it.
- Human-in-the-loop approval before every consequential tool
- A run tree for each delegated task, with its cost summed as it goes
- The title bar shows a pending-approval badge and the open tree’s live cost
Configure
Set it for the org. Override it per agent.
One panel for the settings that belong to the whole organization — the same values wherever you open the app.
- A live model catalogue with tools, no-tools and vision badges, across reasoning, image and speech
- Web access: text only, browser, or both
- Provider keys encrypted at rest by your OS keychain
- Delegation limits: depth, runs per task and a cost ceiling
Under the hood
What runs behind the panels.
Each tool sits on its own rung of the clearance ladder, and every run is supervised, metered and cancellable.
Durable approval
A paused run survives a service restart — a checkpoint plus a run manifest rebuild the graph where it stopped.
Worker–evaluator loop
A second LLM call grades each answer against a success criterion and sends it back with feedback until it passes.
Sequential & parallel delegation
Chain reports one after another, or fan several out at once. Each child runs at its own seat’s clearance.
Budgets that nest
A tree-wide USD ceiling plus a per-agent one over its own subtree. Whichever is lower binds.
Cascading cancel
Stop a root task and every descendant stops with it, deepest-first.
Memory in three scopes
Private to the agent, owned by the seat, or shared by the team — each gated differently, and ownership beats seniority.
Reading the web
fetch_url at clearance 1, behind an SSRF guard that checks the resolved IP at every redirect hop.
A watchable browser
At clearance 5, a real Chromium pane with a persistent session per agent. Approvals read like sentences.
Images & speech
generate_image and generate_speech at clearance 3, written into the workspace and metered into the task budget.
Announcements
A chime or a spoken line when an agent replies or finishes, using your system’s own voice. Free and offline.