The work recordLive
Where the work is, and what it is allowed to cost.
The tasks your agents take, the approvals you give them, and the budgets they run inside. Hosted Paperclip for organisations with a tenant of their own; the AI3 console for everyone else.
Work from Claude or ChatGPT What it costs to run
On your own organisation it is at /tasks.
- Evidence before approval
Every ranked action can be opened to show what it is based on before you say yes to it.
- Caps are enforced, not advisory
Daily and monthly limits are applied at the model provider — the key stops working when the cap is reached — and are never raised automatically.
- Dry run by default
Every write asks for confirmation. Called without it, a tool tells you exactly what it would have done.
Why it exists
An agent you cannot stop is not an employee, it is a liability. This record is the answer to three questions an owner asks constantly and should never have to chase: what is waiting on me, what is it based on, and what can this cost me if I am wrong. Approvals, tasks and budgets sit in one place because they are one decision.
How it works
Work arrives as a ranked action
Your feed is everything waiting on you, most important first, across every organisation you own.
You ask why
The evidence behind an action — the figures, the ledger lines, the thread it came from — before any of it moves.
You decide
Approve, modify, reject, defer or delegate. An approval becomes a task on the organisation, assigned to the responsible agent unless you name another.
It runs, and reports
Finished work goes back through the product manager for review, against what was originally reported. The agent does not tell the reporter it is done; the review does.
The budget holds the line
A cap per portfolio or per organisation, with the month’s usage against it, alerts, and an emergency stop that halts everything at once.
A morning, from one feed
- 09:00Six things are waiting
Ranked across four companies. The top one wants to raise a supplier price by 9%.
- 09:01The evidence
Twelve weeks of orders, the last three invoices from that supplier, and the margin at both prices.
- 09:02Modified, not approved
Six per cent instead of nine, with a note. It becomes a task on the organisation, assigned to the desk that raised it.
- 09:30A cap is reached
One company hits its daily limit. Its key is disabled, the alert is on the feed, and nothing is quietly raised to keep going.
- 14:10Reviewed, then told
The work comes back with what changed and how it was checked. Only then does the reporter hear it is done.
What an agent calls
Over the AI3 MCP server from Claude or ChatGPT, and over HTTP with a personal token. Every write asks for confirmation first; without it, the tool says what it would do.
create_paperclip_taskPut work on an organisation directly: a title, a priority, and who does it.get_task_statusWhere one task has got to.approve_actionApprove a ranked action; it becomes a task assigned to the responsible agent.set_budgetA daily or monthly cap for the portfolio or one company, enforced at the provider.
What it will not do
- Every write needs confirmation. Without it, the tool describes what it would do and does nothing.
- A cap is never raised automatically, by anything, for any reason.
- An agent does not report its own success to whoever asked. The product manager reads the work against the report first.
What it connects to
The systems this record mirrors to or draws on. Each row says whether it is switched on for this server, because a wall of logos that all look connected is the one thing these pages must not be.
- Claude and ChatGPTConnected on this server
AI3 is an MCP server: your organisations, their figures, actions and customers in the conversation.
/mcp - AnthropicConnected on this server
Claude, on your own subscription by default, or a key under Advanced.
Agents & models - OpenAIEach organisation connects its own
ChatGPT via Codex on your subscription, or a key.
Agents & models - OpenRouterConnected on this server
Any model, with Auto as the default when nothing else is set.
Agents & models - PaperclipConnected on this server
The agent runtime hosted per organisation, with the books as a plugin inside it.
Agent operations
Tasks and agents is one of 5.
The core is the infrastructure AI3 builds and runs itself and hands every organisation on the day it exists. Read the rest, or see what your own systems plug into.
Work from Claude or ChatGPT What it costs to run All integrations