ChatGPT cannot see your HubSpot, your Salesforce, your Gmail, or last quarter's closed-won notes.
Every "go look it up in Salesforce, then come back and ask me again" is a broken chain of thought.
Half the questions you ask are ad hoc. You do not want a YAML file for "who bounced this week?".
An LLM that confidently hallucinates revenue numbers is worse than no answer at all.
Mention an agent or just describe the job — the router picks the right one (Outbound, Pipeline Ops, Reply Guy, …) behind the scenes.
Every reply reads from the on-disk Context store — your playbooks, ICPs, past deals — not the open web.
Drop a CSV, a deal record, a LinkedIn URL, a Gong transcript. The chat turns it into structured input for the agent.
Chat writes to HubSpot, fires sequences, runs SQL against your warehouse. Every action is logged with a reversible receipt.
Numbers link back to the CRM record or Context file they came from. One click to verify, one click to edit the source.
Each agent ships with 6 starter chips tuned to its real doctrine — no blank-canvas paralysis.
Do not memorize slugs. Describe the outcome and the router dispatches to the specialist that knows how to deliver it.
Drafts go into Drafts. Sequences start. CRM fields update. Chat is the command line for your GTM motion, not a glorified notes app.
Every thread is saved, starrable, searchable. Promote a good one-off into a starter prompt or a saved agent.
The default is always the latest frontier model. Pin a specific model per-agent when determinism or cost matters.
Upload spreadsheets, PDFs, decks. The agent writes back markdown, CSV, or JSON you can paste straight into a board.
Dictate your prompt between meetings. Whisper-class transcription with your own vocabulary of buyer names.
Long-horizon tasks (10-minute research, 500-row enrichment) run in the background. The composer unlocks instantly.
⌘K to switch agents, ⌘↵ to send, ⌘/ to cite. Built for people who live in the command line.
When a chat thread is something you want to repeat, one click turns it into a named agent with its own prompt and tools.
No. Chat is grounded in your on-disk Context (playbooks, ICPs, past deals) and has real authenticated tools into your CRM, inbox, and warehouse. ChatGPT has neither.
gpt-5.5 by default, as of April 2026. You can pin a different model per-agent or per-chat — Opus, Sonnet, or your own provider via BYOK.
Yes — every reply shows an expandable trace of the tool calls, their arguments, and their results. Nothing is hidden.
The desktop app keeps Chat History and Context local. Model calls need network, but read-only sessions against cached answers work in airplane mode.
