Founder pricing: up to 75% off a Meeteor license · 100 seats per tier, per platformSee seats left →
← All posts

Before, during, after: where a coding agent fits around a call

The last post was about seeding a project with Claude Code: point it at a folder, get a brief, notes, documents and a glossary. This one is about the pattern that fell out of doing that for every call since. A call has three phases, and an agent belongs in two of them.

Before: the agent builds, refreshes and checks

The first time, the agent builds the project from the folder. Every time after that, it refreshes it. A refresh is the same commands with the same folder: ingest skips every file the project already has and adds the new ones, glossary import merges by term name, and the agent replaces the status note with one dated today. Five minutes, mostly spent reading.

Then it checks its own work, and this is the part I did not expect to matter so much. meeteor-cli search runs the same search the overlay runs during a call. So I give the agent the questions I expect to be asked and let it tell me what the project returns for each one. If the uplift cap comes back from the right clause on the right page, good. If a question comes back with nothing, I know before the call, not during it, and the agent usually knows which document is missing.

meeteor-cli ingest --project "Acme Renewal" C:\Clients\Acme\proposal-v4.docx
meeteor-cli note add --project "Acme Renewal" --title "Where things stand (Oct 26)" --file status.md --pin
meeteor-cli search --project "Acme Renewal" "what notice period did we agree"
meeteor-cli ask --project "Acme Renewal" "what is still open before Dana can sign"

During: no agent at all

During the call, nothing orchestrates anything. Meeteor listens, retrieves and shows cards; the CLI is not running and must not be, because it writes to the same database the app is using. This is on purpose. The live path is the product, and it has no tools that act on the world. A card can quote your contract. It cannot send an email, move a deal or change a file, and neither can anything it hears on the call. Documents and transcripts are data to Meeteor, never instructions.

After: Meeteor remembers, the agent acts

When the call ends, Meeteor writes the summary: decisions, commitments, open questions and action items, each linked to the moment it was said. That becomes memory for the next call, and it is what the overlay recalls when the same topic comes up a month later.

Then the agent gets its second turn. meeteor-cli project export writes a project out as readable files: the transcripts, the summaries, the originals and a project.json. An agent can read that folder like any other. Give it the export and it can draft the follow-up email from the commitments, with the exact wording from the transcript next to each one so you can check it, and list the action items with their owners. Where that ends up is up to you and your agent: an email draft, a ticket, a row in whatever tracks your deals. Meeteor does not know or care; it handed over text files.

Before the call, the agent reads so I don't have to. During the call, Meeteor remembers so I don't have to. After the call, the agent writes it up so I don't have to. I still have to take the call.

The rules we give the agent

  • Back up first. meeteor-cli backup before anything is written. One command, one zip.
  • Never while a call is running. Refresh before, export after, nothing in between.
  • Date the notes and replace the stale one in the app rather than adding a second.
  • Rename files on the way in. The file name is what a card cites.
  • Keep a "skipped on purpose" list. Secrets, other people's personal files, the generic playbook. The next refresh reads the list and skips them again.
  • Prove it with search before saying done. Three questions, three passages, or it isn't finished.

What the agent never sees

The CLI and the export folder are the whole interface, and nothing in it touches a key. Building a project needs no AI key at all: parsing, OCR and indexing run on your PC. Your Anthropic or OpenAI key stays in Windows Credential Manager or the macOS Keychain, used by the app for cards and by ask, never written to a file an agent could read. The agent you run does see the documents you point it at, under its own provider's terms. That is a choice you make per folder, and if a folder should not go through an agent, run the same commands yourself.

What might change, and what won't

Could an agent query a project directly, without the export step? Possibly, later; nothing is promised and there is no date. What will not change is the line above: the live path keeps no tools that act, and nothing said on a call can make Meeteor do anything but show you a card. The agent works around the call. Meeteor works inside it.

Founder program

Bring your whole context to the next call.

Founder licenses are a one-time purchase per major version, up to 75% off, and limited to 100 seats per tier on each platform.