Example response · illustrative, not a real run
The AI company operating system
Your company. Powered by AI. Led by you.
Brief your AI chief of staff. It coordinates research, planning and delivery across your departments, then brings the work back for your review.
Free starter planned for early access. Runs locally in Claude Code; your Claude plan or API usage is separate.
One request in. One answer out.
Tell your chief of staff what you need. It brings in the right departments and combines their work into one response.
Illustrative walkthrough · not a live agent run
brief · build a feature5 departments involved
1Start with a brief
Write it the way you would brief a person: what you want, what matters, what must be included. Your chief of staff restates it and works out which departments own the work.
your briefexample
you wrote
2The right departments get involved
The brief goes to the departments that own that kind of work, and only those. Each department head splits it into tasks for its specialists.
who's involvedexample
departments
how many
3Specialists work. Reviewers check.
Specialist agents draft the work against the brief. Before it reaches you, a separate reviewer agent checks it and marks each key claim as taken from your brief, assumed, or needing confirmation. AI review can miss errors, so important claims are listed for you to confirm.
department reportexample
written bya specialist agent in the owning department
checked bya separate reviewer agent, against the brief
resultreviewed · open items listed for you
4You approve consequential actions
Preparing work and taking action are kept apart. Spending, signing, publishing and similar actions pause for your approval; the Approvals section explains which protections are enforced by software and which are workflow rules.
paused actionexample
requestsign vendor agreement
statuspaused · waiting for your approval
5Get the findings and next steps
Everything comes back as one response: a recommendation, the findings behind it, what is still uncertain, the decisions that need you, and the next steps. See an example below.
The work, brought together.
Get a recommendation, the findings behind it, unresolved disagreements and the decisions that need your input.
Reviewed against a clear brief
Each team works to defined requirements, so you can see what was checked and what remains unresolved.
Disagreements stay visible
When departments reach different conclusions, the answer explains the trade-off rather than hiding it.
Tags in the example: from your brief = taken from the brief · assumption = assumed, not checked · to confirm = needs checking before you rely on it
Specialist teams. One point of contact.
Polity brings together 111 AI specialists across 17 departments, coordinated by your chief of staff. Explore the teams and the work each handles.
the team
111
seats · one per AI specialist
Select a department to see what it handles.
How the 111 are counted
82 specialists sit in the 17 departments listed here. The engineering pool, 28 agents coordinated by a lead engineer, is shown hollow and sits outside the 17 so that build work can be routed to it from any department. One separate fact-checking agent, the large centre seat, reviews key claims before a response reaches you. 82 + 28 + 1 = 111, taken from the agent definitions in the repository.
Also on the team:
Work moves forward. Important actions wait for you.
Polity separates preparing work from taking action. Defined checkpoints require review, and protected actions wait for your approval.
1
Payments
Commands that match signing, approval or secret-reading patterns are blocked by a guard script; other spending paths are workflow rules. Blocked by software, plus a pause rule
2
Contracts
Agents draft and review agreements; signing is yours. The workflow pauses at a review checkpoint before anything is signed. Agents are told to pause; not enforced by software
3
Publishing
Connected-tool deploys and outbound messages, and edits to deploy or CI files, are blocked by the guard scripts until you override; a plain shell command is not. The security review before a release is a pause rule. Blocked by software, plus a pause rule
4
Review
A separate reviewer agent checks work before it reaches you, and code is reviewed before it is committed. AI review can miss errors. You remain responsible for reviewing important decisions. Agents are told to pause; not enforced by software
example · a paused action
Request: sign vendor agreement
Review: terms changed since approval
Status: paused
Next step: review the changes and approve again
How approvals work
Blocked by software (3 guard scripts). Three scripts run before an agent's tool call and block it: one stops shell commands that sign, approve, read secret files or start live trading; one stops writes to secrets, keys, .env files, deploy and CI configuration, hook settings, approval settings and the agents' own routing files; one stops connected-tool calls that move money, send messages, deploy, delete, run raw SQL or code, change keys or access, or merge code. Each can be overridden by an environment flag, and the connected-tool guard also has a small hard-coded allowlist; the shell and file guards fail open on a parse error so they never break ordinary commands, while the connected-tool guard fails closed for money and secret servers.
Agents are told to pause; not enforced by software (5 checkpoints). Five checkpoints are written into the agents' instructions: a risk review before anything goes live, a key-and-secret review before key operations, a security review before a deploy, a security review before any auth, payment, wallet or financial code change, and a code review before a commit. These are rules the agents are told to follow, not blocks the software enforces, so a mistaken or manipulated agent could skip them. The guard scripts exist underneath for that reason. Everything an agent does outside these protections relies on your own review.
Try Polity before choosing a plan.
We're starting with a free starter version that runs locally in Claude Code. Try a complete workflow, see whether it helps, and choose whether to upgrade as more capabilities become available.
- 1
Follow for early access
The early-access list is not open yet. When it opens, we'll invite people in small batches.
- 2
Try the free starter
Use a focused selection of specialists coordinated by your chief of staff, with the relevant approval checks included.
- 3
Upgrade if it's useful
Optional paid upgrades are planned to add broader workflows, more specialist teams, maintained updates and support. Details will be confirmed before purchase.
The starter has no Polity fee. Claude Code and your own eligible Claude plan or Anthropic API billing are required. Anthropic's usage limits and charges apply separately.
Questions, answered plainly.
What is Polity?
An AI company operating system: specialist AI agents organised into departments and coordinated by a chief-of-staff agent. You give one brief and get back coordinated work, the findings behind it, the trade-offs, and the decisions that need your input.
Who is it for?
Founders and small software teams who want research, planning and delivery work prepared by specialist agents, with a person keeping the final say. The early-access version expects you to be comfortable working in a terminal.
Is Polity free?
We plan a free starter version with no Polity fee. Optional paid upgrades are intended to add further capabilities, updates and support. Final paid plans have not been announced; anything we publish about pricing is planned, not an offer (Terms §2).
What is included in the starter?
A focused selection of specialists, coordinated through the chief of staff, to demonstrate a complete workflow, with the relevant approval checks included. The exact contents will be confirmed before invitations go out.
Will I have to upgrade?
No. Upgrades are optional. The starter you receive will not be remotely disabled. Compatibility with future Claude Code versions may require updates.
When can I get access?
The early-access list is not open yet; follow @polity_hq to hear when it opens.
What do I need to run it?
Claude Code installed on your own computer, with your own eligible Claude plan or Anthropic API billing. Polity is a set of agent definitions, checks and guard scripts that run inside it. No separate server or installation is needed today.
Does the free version include AI usage?
No. Polity runs inside Claude Code using your own eligible Claude plan or Anthropic API billing. Anthropic's usage limits and charges apply separately, and Polity does not meter them.
Is this a hosted app?
Not currently. Early access runs locally inside Claude Code on your computer; there is no hosted app or web dashboard. Your requests are processed by the model providers and connected tools you use, so running locally does not mean all information stays on your computer. Whether a hosted version follows is not yet decided.
What can agents do without approval?
Read, research, draft, plan and make ordinary changes inside your project. Commands that match signing, approval or secret-reading patterns, writes to protected files, and connected-tool actions that move money, send messages, deploy or delete are blocked by the guard scripts unless you override them. Everything else relies on the workflow rules the agents are given, so you should still review what they did. See Approvals.
Can the agents make mistakes?
Yes. Agents can misread a brief, assume something that is wrong, or produce work that looks finished but is not. A separate reviewer agent checks work before it reaches you and marks what is assumed or unconfirmed, but AI review can miss errors. You remain responsible for reviewing important decisions and anything that is signed, spent or published. See Terms §4.
Where is my data processed?
Your files stay on your computer, but anything an agent reads or writes is sent to the model provider behind Claude Code to be processed, and to any tools you choose to connect. Polity itself runs no server for the product; before the starter is released, we will confirm what, if anything, it sends to us. The website is hosted and served by Vercel; the privacy notice explains what is processed when you visit.
Be first to try Polity.
The free starter version is coming. We'll invite people in small batches once the early-access list opens.
Paid upgrades will be optional. Your Claude plan or API usage is separate.
Early access is not open yet.