Run it · Agent screen
One conversation surface, and it is the screen you land on
The root route is a full-screen conversation. It absorbed the dashboard and the ask screen, because the page you land on should be the one you instruct. The docked copilot that used to float on every other route was removed rather than replaced. Two surfaces for the same conversation is the product arguing with itself about where you type.
Eleven. Nine came from the rev-ops hiring signal and two reacted to a post about attribution. All eleven are in ICP and none has an outbound row yet.
The Series-A fintech ops campaign has capacity today. Want me to enter them?
What is actually going on
Every claim below is followed by the mechanism under it, because a claim without a mechanism is a slide.
The allowlist is the boundary, not the prompt
A model returns a tool name and anything outside the belt built from your own membership is rejected before execution. Injection arriving through a contact record, an agent prompt or an inbound reply can at worst name something that does not exist.
The belt is assembled, never authored
It is the MCP tool list filtered by your membership, plus the capabilities MCP does not cover, plus one navigation tool. Adding a tool in one place adds it in both, so the in-app assistant and an outside client cannot end up with different powers.
A turn is stored as blocks, not as prose
Reopening a thread shows the sentence and the reads that produced it, in the order they happened, because the sentence is a claim and the tool calls are the evidence for it.
Nothing on the belt sends
The launcher and its gate remain the only path to a person. A screen that could queue or send would put a model between your sentence and somebody's inbox.
Step by step
How the agent screen runs
- 1
Ask
Plain language, or a slash command to address one role directly.
- 2
Plan
The model picks tools from the belt your membership allows.
- 3
Run
Each call is checked again at execution. The advertised list is a convenience, the second check is the boundary.
- 4
Answer
In its own voice, with the tool calls kept beside it as the record.
The numbers
Facts, not adjectives
Every figure here is a real value from the running product. A cadence from the scheduler, a price from the credit table, a limit from the ramp. A marketing page that invents a limit is a support ticket with a delay on it.
- Belt source
- MCP tools + capabilities + navigation
- Scope check
- Twice. On listing and on call
- Metered as
- ai.turn.standard per model pass
- Confirmations
- None. The machinery stays for the next irreversible tool
- Transcript
- Full blocks, tool output included
The agent screen, honestly
Does it ask before it spends?
No. Metered spend runs unconfirmed, governed by the credit ledger and by your own membership rather than by a chip. That is a deliberate decision, not an oversight.
What happens if I close the tab mid-turn?
A replayed running tool is shown as errored rather than as a spinner that spins for ever. Turns that outlive the browser are a build in progress, not a claim we make today.
Run it
The team that does all of the above
The agent screen is one part of it. The rest runs too.
Sourcing, enrichment, intent, copy, sending, replies and reporting are one product with one bill and one gate. Connect a channel and the whole motion starts.
Start with sales. $2,000 or $5,000 onboarding, then at least $250 monthly credit funding.