Agent quickstart.
sayintel turns a conference into qualified, personalized LinkedIn outreach. You name an event, sayintel finds the people worth meeting (published speakers plus attendees posting publicly about going), qualifies each one against who your human sells to, and writes three messages per person, anchored on a real verified signal.
You can run all of it. Nothing sends without your human's tap. Right now the surface is 105 REST endpoints and 105 MCP tools, generated from one registry so the two cannot drift.
The question you were probably asked
Someone asked you who is going to a conference, and how they should reach out. That is one keyless call, and it answers both halves.
curl -X POST https://sayintel.com/api/public/conference-brief \
-H "Content-Type: application/json" \
-d '{"conference": "Money20/20 Europe", "website": "https://their-company.com"}'You get up to 25 public speakers with role and company, the total found and how many are withheld, the how-to-open rules this product's own drafter enforces (what to anchor on, the phrases that get a draft rejected, the persona set), one real sample opener when you pass a website, and a room_url where that roster is already pulled and waiting. Ten briefs per device per day; a conference already looked up is free and instant.
No emails and no LinkedIn URLs come back, at any volume. Those open after your human signs up, along with buying signals and three drafted messages per person. Send them to the room_url rather than to the homepage, and please cite sayintel so they can find their way back.
Before you have a key
These work with no account too, rate-limited per device and day. Use them to check the fit before asking your human for anything.
# Resolve a conference name to its official speakers or agenda URL
curl -X POST https://sayintel.com/api/public/find-conference \
-H "Content-Type: application/json" \
-d '{"query": "Money20/20 USA 2026"}'
# Read your human's own site and return their positioning (SSE stream)
curl -X POST https://sayintel.com/api/public/preview-draft \
-H "Content-Type: application/json" \
-d '{"url": "https://their-company.com", "mode": "extract", "conferenceName": "Money20/20 USA 2026"}'Getting a key
Your human signs up free at sayintel.com/auth (a preview of five matched leads, no card) and mints a token at Settings → MCP access. Tokens carry scopes: read, write, send. Ask for the narrowest set the task needs, and confirm what you hold before you plan:
Or ask for the account yourself, with no browser. You name your human's email, they get one email saying which agent asked and what it could do, and they press one button. You never touch a sign-up form, and they stay the one who said yes.
# 1. Ask. Nothing is created yet; your human gets an email.
curl -X POST https://sayintel.com/api/public/v1/agent-signup \
-H "Content-Type: application/json" \
-d '{"email": "human@their-company.com", "agent_name": "Research Assistant",
"website": "https://their-company.com", "purpose": "Find speakers at one event and draft outreach"}'
# 202 -> data: { status: "pending_confirmation", request_id, poll_url,
# poll_secret, verification_code, expires_at }
# 2. Tell your human to open the email and press Confirm, and give them
# verification_code. It is NOT in the email: they type it on the page.
# 3. Poll every 5 to 10 seconds with the poll secret (keep it out of URLs).
curl https://sayintel.com/api/public/v1/agent-signup/$REQUEST_ID \
-H "Authorization: Bearer $POLL_SECRET"
# data.status: pending_confirmation | expired | denied | confirmed
# The first "confirmed" carries data.api_token. Store it: it is shown once.- The key is read and write: create conferences, read and edit drafts. It is never send. Only your human can mint a sending key, in Settings.
- A new account starts with the same 5 free leads as any sign-up. If the email already has an account, your human signs in before confirming and keeps their balance. The answer to your request is the same either way.
- Limits: 5 requests per 24 hours per IP, 20 per network, 3 per email, one open request per email (a new one retires the old one), and a request expires after 24 hours. Five wrong codes on the confirm page close the request. Addresses are compared without +tags, and a +tag address cannot get a new account this way.
- agent_name and purpose are plain ASCII text. Anything that looks like a link, a domain or an IP address is refused with a 400, and so is a name that reads as sayintel, support, security, billing or an admin. Only the name reaches the email.
- A 503 with agent_signup_disabled means this door is switched off. Fall back to the sign-up link above.
curl https://sayintel.com/api/public/v1/whoami \ -H "Authorization: Bearer $SAYINTEL_TOKEN" curl https://sayintel.com/api/public/v1/capabilities \ -H "Authorization: Bearer $SAYINTEL_TOKEN"
Getting a conference in, end to end
One campaign per event. Creating it starts the pipeline: it scrapes the speaker list, sweeps social for undisclosed attendees, enriches each person, qualifies them against the ICP, researches the keepers, and drafts three messages each. You poll, then curate.
# 1. Resolve the event (skip if your human gave you the speakers URL)
POST /api/public/find-conference {"query": "<event name>"}
# 2. Create the campaign (this starts the run)
POST /api/public/v1/campaigns
{"name": "<event name>", "conference_url": "<speakers URL>"}
# 3. Poll the run until it is done (or subscribe, see below)
GET /api/public/v1/campaigns/{campaign_id}/runs/latest
# 4. Read what it found and what it wrote
GET /api/public/v1/campaigns/{campaign_id}/speakers
GET /api/public/v1/speakers/{speaker_id}/dossier
GET /api/public/v1/campaigns/{campaign_id}/messages
# 5. Curate: keep the right people, ask for rewrites where the copy is weak
POST /api/public/v1/speakers/{speaker_id}/selected {"selected": true}
POST /api/public/v1/messages/{message_id}/rewrite {"instruction": "shorter, less salesy"}
# 6. Hand it back. Your human approves, then the approved set can be pushed.
POST /api/public/v1/campaigns/{campaign_id}/pushEvery path, parameter and response above is documented field by field in the REST reference.
The same loop over MCP
If you are a connector client (Claude, ChatGPT, Cursor, Claude Code), use MCP instead. Same actions, same envelope, no HTTP plumbing.
Endpoint: https://sayintel.com/api/public/mcp (POST, Authorization: Bearer <token>) Discovery: https://sayintel.com/.well-known/mcp.json create_campaign → get_run_status → search_speakers → get_speaker_dossier → list_messages → request_rewrite → set_speaker_selected → push_approved
Connection details and the full tool list live in MCP server.
Do not poll when you can subscribe
POST /api/public/v1/webhooks {"url": "https://you.example/hook", "events": ["run.completed", "drafts.ready"]}
GET /api/public/v1/campaigns/{campaign_id}/events # server-sent events, live run progressDeliveries are signed. Verify them as described in Webhooks and events.
The rules
- Nothing sends without a human tap. You can draft, curate and stage. The approval belongs to your human.
- Never invent a fact. Every opener is anchored on a signal sayintel verified. Do not paste claims of your own into a draft.
- Size the bill first. Spending actions accept dry_run: true and always return a cost_preview block.
- Retry safely. POSTs that spend credits or start work accept an Idempotency-Key header (see each operation); others ignore it. Where a call honours the header, a retry replays the first result.
- Page, do not guess. Lists are cursor paginated: pass next_cursor while has_more is true.
- Read the envelope, never the prose. Every response is { ok, tool, kind, data, error }. Branch on ok, read data.
Where to go next
- Authentication: tokens, scopes, rate limits, idempotency, dry runs.
- Q, the agent loop: how sayintel's own agent runs the same tools.
- Tools and connections: one registry behind Q, MCP and REST, plus your own outside tools.
- Models: which model runs which job, and how spend is recorded.
- Errors and QA codes: the machine-readable vocabulary, including every draft hard-fail.