I wanted the "who just followed which new AI startup" signal to land in my own scripts, not just in a browser tab. Notes from wiring that up, using Earlyfollow's MCP server and HTTP API as the example.
What counts as a new follow
Each day the job re-reads the follow lists of the tracked accounts and keeps only the accounts that appear ahead of the ones already followed. Position, not timestamps, because X does not publish when a follow happened — the date you see is the day the follow was spotted. Follows that existed when tracking started are kept as a baseline and never presented as new. That kills most of the noise from unfollows and reshuffles.
Connect over MCP
For Codex, two lines:
codex mcp add earlyfollow --url https://earlyfollow.com/mcp
codex mcp login earlyfollow
Any other client that speaks HTTP transport can register the same URL; the auth step follows the client's own OAuth flow.
Call the HTTP API
For a cron job, skip MCP and POST straight to the list endpoint:
curl -s https://earlyfollow.com/api/v1/list-signals \
-H 'content-type: application/json' \
-H "authorization: Bearer $EARLYFOLLOW_API_KEY" \
-d '{
"request_id": "daily-2026-11-21",
"window": 7,
"sort": "concentration",
"kind": "company",
"query": "agent"
}'
A project API key is required; without one the response is authentication_required, and over MCP the codex mcp login session covers it.
Parameters are narrow on purpose: window is 7 or 30, sort is concentration (how many tracked accounts follow it) or latest (detection day), kind is company or person. Personal accounts only show up on the paid side, so ask for kind=company when you want something close to the public list.
Use request_id as an idempotency key
The quiet win is request_id: it must match ^[-\w:.]{1,120}$, and results are retained for 30 days. If a response gets lost, replaying the same ID returns the same result instead of fetching twice. For a daily job, "run date + window" is a good ID shape — a retry then just re-runs the same call.
When you aggregate it
Most new follows land on people, not companies. To present a company list you also need to resolve the profile to a company and keep the rest as people. And a follow is a record of attention, not a forecast. Public pages and the free weekly list run 14 days behind, so anything that needs the last few days has to come from the paid side.
If you want to see the same data as a human-readable board, that lives at https://earlyfollow.com/.
Top comments (0)