Bring your coding agents into one workspace. Choose the model for each role and follow the team while they work.
Everything the team does stays readable, and every decision stays yours.
Directors see the whole task, managers split it, workers execute focused assignments. Each agent works in its own context window; concise handoffs carry decisions between them.
You set the goal and can see what is happening at every step. Nothing runs, merges or ships without a visible decision.
Assign the right model to each role by capability and cost. Change your mind per project without changing tools.
Models from different providers work as one team. A GPT director sets the goal, a Claude manager splits it, a Claude senior and a Kimi junior build it, and they review, question and correct each other in the same conversation you are reading. No glue code, no copy-pasting between tools.
Spend where the thinking matters. Give each role the model that fits its job and budget: a capable model for scoping and review, lower-cost models for bounded work. Swap a model on one role and the rest of the team keeps working. The model name sits beside every message, so you always know who said what, with what.
Provider usage shows what is left on every account you connect, across all projects: Codex weekly windows, Claude's five-hour and weekly windows, your OpenRouter balance, each with its reset time. Bars show what remains and refresh every minute.
When a window runs low, switch that role to another model or provider and the team keeps going. Claude's bar shows the last reading Claude itself reported; Agency never reads your login to get it.
A project is more than code. Agency comes with ten departments that follow the shape of a real venture, from the first idea through market research, a business plan and product specs to architecture and active development. Open the ones you need; each has its own manager, its own channel and its own workers, and they run at the same time.
Every teammate shows the provider and model it runs on and how long it has been working. All updates gathers every department into one stream, so you can follow the whole project from a single place.
Not a pipeline. The director runs several departments at once, sends work back to an earlier stage when the build calls for it, and keeps every channel in one stream.
A goal is a sentence, not a ticket. These are the kinds of things people point a team at. The director reads the goal, opens the departments it needs and staffs them — and comes back to you when a decision is yours to make.
“Compare how our five closest competitors price their team plans, and tell me where we look expensive.”
Two workers are splitting the list. You will get one table with every number linked to the page it came from, and the three plans where we are the outlier.
“People drop out on the second onboarding screen. Give me two ways to fix it.”
Two flows drafted: one removes the step, one splits it in half. Both written up with what they cost to build. The second needs your call on what we ask for at signup.
“Work out what we would have to charge to break even at five hundred customers.”
Model built from the cost list in the project folder. One input is missing — support cost per customer — so it is left as a variable rather than a number I made up.
“Draft the launch note and three short posts, in our voice, for the Windows release.”
Draft and three posts, following the tone guide in the project folder. The pricing line is deliberately blank: give me the number and I will fill it in everywhere.
“Three customers report the same failed checkout since Tuesday. Find out why and fix it.”
Reproduced from the reports: a saved card plus an expired session. The fix is two lines and a regression test. Before it goes out, say whether the three orders get refunded.
“We have three weeks and one engineer. Tell me what is worth building next and what to drop.”
Eleven items scored against the three weeks. Four fit, and two of those are things nobody asked for. One ranking turns on whether the enterprise deal is real — that one is yours to confirm.
Every project gets its own team, its own conversations and its own terminals. Open as many as you are working on and switch between them without stopping anyone.
When a team needs a decision, the project says so from the list, so you can leave one team building while you answer another. Nothing is shared between projects unless you share it.
Turn on Auto check-in and Agency asks your director to review the project goal on a schedule, every one to five hours, and to carry on with whatever is still open. If no director is on duty, it starts one.
Every check-in leaves a record: goal reached, more work remaining or not reached yet, plus the tasks that need you. When a goal is reached, the director is asked to propose the next one. The schedule is a timer, not a model; the thinking is done by your team. It runs on your Mac while Agency is open, and you switch it off whenever you like.
The Looper asks your director to review the project goal on a schedule and continue any outstanding work.
The readable conversation has a real execution layer underneath. Switch any agent to its terminal, watch the actual commands and output, then return to the conversation where you left it.
Agents run on your machine, with the providers you already use. Selecting a project folder is you trusting it; review overrides live in Settings.
$ npm test -- provider-routing > agency@[version] test > vitest run src/shared/llmProvider.test.ts ✓ src/shared/llmProvider.test.ts (14 tests) 812ms Test Files 1 passed (1) Tests 14 passed (14) $ git switch -c feat/provider-routing-tests $ git commit -am "test(provider): routing suite on the mock provider" $
Claude and Codex connect through their native sign-in on your machine. OpenRouter models use an API key you add in Settings. Ollama runs local models with no account. Existing eligible accounts can be reused; Agency does not pool or resell subscriptions.
On your computer, in the project folder you choose. Agency starts each agent's session locally and shows you its terminal. Cloud execution is not part of the current release.
Providers bill their own usage under your accounts. Agency shows the model beside every message and keeps per-role model choices visible, so you can decide where the stronger model is worth it. It does not promise savings or pick the cheapest model for you.
No. Each agent works in its own context window and receives a focused assignment. Managers coordinate through concise messages and handoff files in the project. A separate context is not a security sandbox; selecting a folder trusts it.
macOS on Apple silicon, and Windows 10 or 11 on 64-bit. The Windows installer is unsigned and has not been run on Windows hardware yet. Linux comes next. Phone apps to follow your team are planned.
Agency coordinates AI agents under one person's direction. Sharing a workspace between people is future work and not part of the current release.
One download per platform. The button at the top of the page already picked yours.
[macOS minimum version]. Apple silicon build. Not yet Apple-notarized.