How Jot works · a blueprint you can copy
Speak it.
Agents do it.
Jot remembers it.
One voice request becomes real work on your own computer. The context stays for the next conversation.
You speak“Put the winter hours on the website.”
ChatGPTunderstands and calls a tool
Jot serverjob queue + notes
Agent on your computerdoes the work, reports back
The whole system on one page
Three places, five parts
In your pocket
ChatGPT applistens, understands, calls a tool, speaks the answer
Resultsask by voice, or see them on the phone
tool call (MCP)
reply
In the cloud
Jot tool serveroffers the tools, checks the sign-in
Job queueevery job: brief, state, report, files
Notes vaultprojects, notes, decisions, history
asks for work
results
On your computer
Runnertakes the next job, gives it its own folder
AI agentdoes the work, writes a report
Claude CodeCodexGemini CLI
Phone ⇄ server: seconds, while you talk
Server ⇄ computer: minutes to hours, in the background
The key idea
A tool server is the bridge
ChatGPTin OpenAI's cloud. It cannot see your computer.
jobs + notes (MCP)
results
Tool servera mailbox that both sides can reach
asks for jobs
sends results
Your computeragents, files, logins
What ChatGPT sees: a tool menu
Jobs
add_jobget_joblist_jobsMemory
write_noteread_notelist_notesProjects
list_projectsproject_overviewMCP (Model Context Protocol) is the standard plug. Each tool has a name, a plain description and fields. ChatGPT picks the tool from the description.
Why this shape
Safe: your computer only calls out
Reliable: jobs wait if the computer is off
One record of everything
Swap ChatGPT or the agent
The server does not think, and it does not do the work. It stores and passes on. So it stays small.
Idea 1 · Voice first
One spoken request, from start to finish
time
Youvoice
ChatGPTunderstands
Jot serverqueue + notes
Your computerrunner + agent
08:10
You → ChatGPT“Put our winter hours on the website. And from now on, we close on Mondays.”
ChatGPT → Jot
project_overview reads the website project and its last jobs08:11
ChatGPT → Jot
write_note saves the decision: “closed on Mondays”ChatGPT → Jot
add_job website, code, clear brief → “Job 57 queued”ChatGPT → You“Job 57 is queued. I saved the Monday rule.”
The call ends. The rest runs without you.
08:12
Computer → Jot“Any work?” · gets job 57
08:40
On your computerOwn branch + folder · agent edits the page · checks · commit
08:41
Computer → Jotreport + screenshot · “done”
12:30
You → ChatGPT“How did the website job go?”
ChatGPT → Jot
get_job(57) → the reportChatGPT → You“Done. The page shows the winter hours and Monday closed.”
Idea 2 · A second brain
Every chat starts with what Jot remembers
Jot vaultprojects · notes
job history
job history
1 Capturesay an idea, a decision, a file
2 Organizeunder its project
3 Delegatea job gets the notes it needs
4 Remembereach report becomes a note
5 Retrieveany new chat reads it back
↺ back to 1 Capture
Kept in Jot
- Project pages and notes
- Decisions and corrections on jobs
- Every job: brief, report, files
- Project rules that each agent gets
- Old versions of each note
Only in ChatGPT
- The current conversation
- ChatGPT's own memory. Agents never see it.
- Anything that no tool call saved
Rule: if it matters later, save it to Jot. A new chat reads it back with a tool call.
Idea 3 · Code jobs
Each code job gets its own branch and folder
Local commitin the job's own branch, on your computerevery code job with changes
Jot uploadreport, screenshots, files to the serverevery job
GitHub pushthe merged main, not the job branchonly at merge, only with a GitHub remote
Build your own
The map: buy two parts, build three
Voice assistant BUYChatGPT with a custom connector
Tool server BUILDsmall web service, official MCP library, sign-in
Store BUILDone job table + a notes folder with history
Runner BUILDa loop on a computer that stays on
AI agents BUYClaude Code, Codex or Gemini CLI, no chat window
Build order
1Jobs:
add_job, get_job2Runner + one agent, one folder per job
3Results back: ask “how did it go?”
4Memory:
write_note, read_note5Code: branch per job, checks, merge
The whole idea: a voice front door, a shared memory and queue in the cloud, and workers on your own computer.