Skip to content
Feature · AI agents

Give your agents a memory.

Earlyn gives AI coding agents a memory: one click registers a local MCP server in Claude Code, Codex, Cursor, Windsurf or Gemini CLI, so they can search what you saw and heard and ask Earlyn questions, over 127.0.0.1 only.

Updated Read as Markdown

Connect in one click

Settings → Connect lists Claude Code, Codex, Cursor, Windsurf and Gemini CLI. Connect writes the Earlyn server into that tool's own settings and turns on Earlyn's local API; disconnecting the last tool turns it off again. By hand it is one command:

claude mcp add --scope user earlyn -- "/Applications/Earlyn.app/Contents/MacOS/earlyn-mcp"

What the agent can do

ToolWhat it returns
search_memoryMoments, transcript lines and documents that match
recent_memoriesThe newest moments, optionally for one app
get_memory / get_frameThe full text or the picture of one moment
list_meetings / get_meetingMeetings with transcript, summary and action items
ask_earlynAn answer from your history with its sources
daily_digests / user_profileDaily summaries and the facts Earlyn learned
query_sqlRead-only SQL over the database
list_skills / get_skill / run_workflowYour skills and agents

Let the agent check its own work

Turn on Let AI tools read the screen in Settings → Connect, and the agent can look at what is on screen now after it changes your code: the running app, the simulator, a browser tab or a terminal. Earlyn answers with words instead of screenshots, so a check costs a fraction of the tokens.

ToolWhat it returnsTypical size
screen_errorsOnly the lines that look like errors or warningsabout 100 bytes
screen_changesOnly what changed since the agent last looked at that appabout 250 bytes
find_on_screenWhether a word, button or message is visible, and wherea few bytes
read_screenThe visible text, top to bottomunder 1 KB
wait_for_textWaits until a word appears, such as a build finishingone call instead of polling
see_screenA picture, only when layout or images matter, cropped if you likeabout 80 KB
similar_errorsWhether you saw this error before and the windows you went through to fix ita few lines
project_contextWhat you did on this repository lately: files, pages read, meetings, open errorsa few hundred tokens
open_loopsPromises, decisions and customer requests from your meetingsa few lines
Measured on an M2 Mac: about 0.14 s per check. The ready-made skill “Check my work on screen” teaches the loop: run, check errors, look at the changes, fix.

Chat that drafts what to send

Ask Earlyn's Chat to do something, like “summarize last week, update Linear and prepare the weekly update for the client”. It answers with real numbers from your memory (time per app and per project, your meetings), then drafts up to three actions as cards: send to a connected app, open an e-mail in Mail, or copy. You can edit each card, and nothing is sent until you press Send.

Skills, agents and the command line

  • Skills Market: reviewed SKILL.md files such as /distill-day, /summarize-meeting and /standup that teach Claude Code and Codex to use your memory. They are bundled in the app.
  • Agents: prompts that run on a schedule, when a meeting ends, when a word appears on screen or when an app opens, and save to your notes or send to Slack, Notion, Linear, GitHub or HubSpot.
  • Command line: earlyn search, earlyn ask and earlyn context give a terminal the same memory.
  • Local API: REST and live events on 127.0.0.1 with a bearer token. See the docs.

Questions

Is the MCP server reachable from my network?
No. The API listens on 127.0.0.1 only and refuses requests without its token. It is off until you connect a tool or turn it on.
Does the agent see everything?
It sees what Earlyn kept, through the tools above, and the screen right now only if you turn on Let AI tools read the screen. Excluded apps, private windows and Earlyn itself are never shown, and a screen with a key or password gives only its redacted text.
Does Earlyn send my screen to Anthropic or OpenAI?
Earlyn sends nothing. Your agent reads through Earlyn on your Mac, and what the agent reads goes to its own service, exactly like text you paste into it.