The principle: Cosa joins your stack, your stack doesn't move

Cosa is an intelligence layer, not a destination system. The design goal is that nobody on the sales team changes how they work to feed it: the data your deals already produce flows in automatically, and the intelligence flows back out to wherever your team and your other tools already live. Two mechanisms deliver that — native connectors for data in, and an MCP server for intelligence out.

Native connectors: what syncs

Five connectors ship today. What each one pulls:

  • Close — calls, emails, meetings, notes, SMS, and tasks across your active opportunities, including full email and call bodies (not just metadata). Authenticated with your Close API key.
  • HubSpot — calls, emails, and deals. Authenticated with a pasted OAuth token.
  • Gmail — email threads, via the standard Google OAuth consent flow.
  • Granola — meeting notes and transcripts, via API key.
  • Google Meet — meeting transcripts, pulled from Google Drive via Google OAuth.

Mechanically, every sync follows the same loop: pull new activity since the last sync, de-duplicate against what was already ingested, and run each new item through the full intelligence pipeline — so a freshly synced call transcript becomes gate assessments, stakeholder updates, and provenance-tagged evidence on the deal, not a file in a folder. Credentials are encrypted at rest (AES-256-GCM), and the Google integrations never see your password — OAuth tokens only.

You can also bypass connectors entirely and paste a transcript directly; the pipeline treats it identically. That is the fastest path on day one — see getting started with Cosa.

The MCP server: 19 tools for any AI in your stack

MCP (Model Context Protocol) is an open standard that lets AI tools call other systems — think of it as an API designed for AI agents. Cosa ships an MCP server exposing 19 tools, which means any MCP-capable AI in your stack — a CRM copilot, an email assistant, a desktop agent like Claude or Cursor — can query your deal intelligence directly, with the provenance tags intact.

What the tools cover, grouped by job:

  • Deal state: full deal overview, Decision Gate progression with evidence quotes, stakeholder map with roles and sentiment, pipeline-wide deal list.
  • Next steps: recommended next action, structured call plans, champion kit generation (exec summary, ROI one-pager, talking points, value defense).
  • Trust and verification: verify a specific metric's provenance, pull the full provenance chain for a deal's data points, generate an honesty report with confidence grading.
  • Pipeline analysis: cross-deal risk radar, pattern correlations, historical deal snapshots.
  • Operations: ingest a transcript, trigger a connector sync, update deal fields, record interaction outcomes, chat with or run Cosa's specialized agents.

The practical effect: your deal intelligence stops being locked in one app. An email assistant can pull the verified metrics for a follow-up instead of guessing. A manager's AI can ask "which deals are single-threaded?" and get an evidence-backed answer. Every tool answers from the same provenance-tagged source of truth — the same per-claim evidence chain a human sees.

CRM write-back: the honest answer

This is the question prospects ask most, so here is the precise current state.

What works today: the read path is fully live. All five connectors pull data in automatically, and any AI tool in your stack can read deal intelligence out through the MCP server. Deal fields and outcomes can be updated in Cosa itself — by you or by your agents through MCP — and generated outputs (briefs, emails, champion kits) are delivered as ready-to-use drafts and exportable files you execute through your own tools, with you confirming each action.

What is not live yet: Cosa does not currently write records back into your CRM automatically. Write-back is in active development, starting with Close (pushing intelligence as notes on the lead), with bidirectional sync on the roadmap behind it.

We state this plainly because the gap between "integrates with your CRM" on a pricing page and what actually syncs is exactly the kind of claim we built provenance to kill. If automatic CRM write-back is a hard requirement for your rollout, ask us where it stands when you talk to us — the answer will be current and specific, and the same honesty applies on pricing.

What this means for an evaluation

Three questions worth testing live rather than taking on faith — ours or anyone's:

  1. 1.Does data flow in without the team doing anything? Connect one source and watch a real call become deal intelligence without a manual upload.
  2. 2.Can your other AI tools use the output? Point an MCP-capable assistant at the server and query a real deal.
  3. 3.Is every synced claim still traceable? Pick a number in the output and follow its provenance chain back to the source line.

A stack integration is only as good as the trust in the data moving through it. That is the actual reason the connectors and the MCP server exist on top of a provenance architecture: integration moves data, but Provenance is what makes the moved data worth acting on.