8000 GitHub - tduda/familiar: An AI chief of staff for product managers and founders. Runs locally in Claude or ChatGPT (Codex): one morning brief, drafts you approve. · GitHub
[go: up one dir, main page]

Skip to content
tdudaPublic

About

An AI chief of staff for product managers and founders. Runs locally in Claude or ChatGPT (Codex): one morning brief, drafts you approve.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Familiar

An AI chief of staff for product managers and founders. Every weekday morning it reads the tools you connected, works out what's due, drafts what it can, and hands you one brief: what changed, what needs you, and what's ready for your yes.

It runs on your computer, in the AI app you already pay for. Free and open source (MIT).

The rules it keeps

  • It runs locally, with your own AI account. Claude or ChatGPT, on your Mac or Windows PC. There is no Familiar server.
  • Nobody else sees your data. Not the person who made Familiar, not anyone. Nothing phones home: no analytics, no telemetry, no update check that sends anything.
  • It drafts, you approve. Every outgoing message, task, post or page needs your yes for that one item. Connectors start read-only.
  • You stay at the keys. Setup is something you do, with the assistant explaining each step. You type every login yourself.
  • It tells you how sure it is. Every claim in the brief carries a marker: verified (re-checked by a second agent), reported (read from one source), inferred (its own reasoning, so check it), or carried (from an earlier day). A number it couldn't read says "not checked".

How it keeps your computer safe

The morning run happens with nobody watching, and it reads text strangers can write (your inbox, chat and calendar). So the limits are enforced by your AI client's own permission system, not just by instructions:

  • Claude, terminal runner: the run may write only inside your workspace, may run only Familiar's own scripts, and may call only the connector tools you list by name (none by default). It can't edit its own config, the engine or the instruction files.
  • Codex, terminal runner: the run is sandboxed to your workspace, its shell has no network, and any connector call that needs approval is refused rather than waiting.
  • OpenCode allows everything by default, so setup writes a locked-down opencode.json, and the runner refuses to start without it.
  • Desktop apps (Claude or ChatGPT): a scheduled task uses the permissions you grant it. Setup tells you to always-allow only the engine scripts and your connectors' read tools; everything else keeps asking, and an unattended ask fails instead of acting.
  • The scheduled runner lives outside your workspace, so nothing the run writes can change what runs tomorrow.
  • Your files stay private: familiar/ is readable by your account only, and nothing is sent anywhere except what you approve.

Details per client: docs/SCHEDULE.md. To remove everything: "Remove Familiar" at the end of that page.

Start here

  • Never used a terminal? Follow docs/GETTING-STARTED.md. Plain steps, the desktop apps first, your first brief in under an hour.
  • Comfortable with a terminal? Clone the repo, run python3 scripts/quickstart.py (Windows: py scripts\quickstart.py), open the folder it sets up, and say /onboard (Claude) or $onboard (Codex).

Want help setting it up? I run guided setup sessions: you stay at the keys, I walk you through it, and you leave with your first brief scheduled and your first jobs handed over. Book a session.

Where it works

Claude desktop app (Code tab) ChatGPT desktop app (Codex) Cursor / VS Code Terminal
Start setup say "set me up", or /onboard say "set me up", or @onboard /onboard (Claude extension) or $onboard (Codex extension) claude then /onboard, or codex then $onboard
Daily brief and drafts ✓ ✓ ✓ ✓
Second-agent verification ✓ checks run inline, so claims are at most "reported" as the extension's agent as the client
Schedule it Code tab > Scheduled, "Only on this computer" ask in a chat: "every weekday at 07:52, run $morning" use the app or the terminal launchd, Task Scheduler or cron/systemd
Runs when the app is open and the computer awake the app is open and the computer awake the computer is awake
Missed run one catch-up run when the computer wakes not documented; Familiar's own check catches it the runner retries until 12:00

OpenCode works too (/onboard), with the terminal runner.

Mac and Windows

Mac Windows
Python python.org installer python.org installer, tick "Add python.exe to PATH"
Start file Start here - Mac.command Start here - Windows.bat
Keep awake while plugged in the app's keep-awake setting, or System Settings > Battery the app's keep-awake setting, or Settings > System > Power
Open the app at login System Settings > General > Login Items Settings > Apps > Startup
Terminal scheduling launchd (docs/SCHEDULE.md) Task Scheduler (docs/SCHEDULE.md)
Brief notification macOS notification and a local page a local page (and a toast from the terminal runner)

Windows support is written and rehearsed, not yet run on many real machines. Tell us what happens on yours.

Setup, in stages

Each stage says in one line what it does, asks only what it needs, and ends with something you can see. Optional stages can be skipped; the brief brings one up later only when it would help with something in front of you.

Stage Time You get
0 · Install 10 min a working setup, checked end to end
1 · About you 15 min a brief built on your goals, and your first job to hand over
2 · Task tracker (optional) 10 min what's yours, what's stuck, what changed
3 · Calendar and meetings (optional) 10 min prep for today's meetings; actions from yesterday's notes
4 · Inbox and chat (optional) 10 min who's waiting on you, with drafted replies
5 · Make it automatic (required) 20 min the brief makes itself every weekday
6 · Your first brief (required) 25 min you read it, approve or skip, and tell it what to change
7 · Hand over a job (optional) 20 min one recurring job drafted for you
8 · Numbers (optional) 10 min your metrics and OKR averages in the brief
9 · Your first week (required) 10 min three more jobs to hand over, about one a day

Stages 2 to 4 connect only what your first job needs. Two ways to do it: alone (/onboard), or guided (/onboard guided), with someone walking you through it on a call while you do every step.

Features you can switch on

Every optional feature starts off, so your first week stays simple. At the one-week review you're offered up to two that fit your week; after that the brief offers at most one at a time (two a week at most), only when it would help with something in front of you. Each offer is fitted to you: it names your own recurring meetings and the reports you already send, and proposes how the feature would work for those. "Not now" and "never" are respected, and "what else can you do?" always shows the full list. Each has a page in features/: what it does, when to switch it on, how, how to get the most out of it, and how to switch it off.

Feature
Meeting prep a prep note before the meetings that matter
Inbox and chat triage who's waiting on you, with drafted replies you approve
Daily numbers any metric, plus OKR 7-day and 30-day averages
Experiment tracking running experiments to their decision date, with a memo on the day
Stall check stuck work turned into one next step
Weekly update your update drafted from what actually happened
Customer intelligence what customers keep saying, by theme, in one living brief (optional)
Research digest up to seven market and competitor changes a week that touch your goals (optional)
Roadmap review your roadmap checked against goals, delivery and evidence; proposes, never edits (optional)
A brief that grows add, sharpen and drop sections by saying so
Threads work the brief in the same chat or a new one
Sprint pack sprint report, planning, metrics, backlog, retro, refinement (optional)
Thinking tools hypothesis, challenge, idea forge, strategy critique, interviews, design review (optional)
Style pass a quiet last edit so drafts sound like you, with your facts untouched
Writing-rules hook your hard writing rules checked before a draft is saved (Claude only)
Skill miner the requests you keep typing, found in your own local sessions and offered as skills you choose (optional)
Ship triage all of today's drafts re-checked at once, one table, a yes per row (optional)

Repo map

commands/    /onboard, /morning, /brief, /tickets, /shipped, /coach (skills in Codex)
stages/      the onboarding stages
features/    the switch-on pages
skills/      the jobs Familiar can draft, plus the optional packs
rubrics/     the quality bar each drafter is graded against
roles/       presets that shape the brief (product manager, founder, and more)
connectors/  one guide per tool, with a test read
agents/      read-only verifier agents (Claude)
hooks/       optional session-start freshness block (Claude)
scripts/     install, setup, doctor, onboarding map, the engine, schedulers
templates/   the files setup creates in your folder
docs/        GETTING-STARTED, SCHEDULE, LAYOUT

Status

Beta. The engine comes from a loop run every working day since mid-2026; this public version is new. tests/run.py rehearses install for every client, setup, onboarding, the engine and every runner with stand-in clients. Not yet tested on real installs: scheduled runs in the Claude and ChatGPT desktop apps with nobody at the computer, the Codex and OpenCode runners, and Windows. Please open an issue with what happened on yours.

Contributing

Issues and pull requests are welcome: read CONTRIBUTING.md first. Found a security problem? Report it privately, as SECURITY.md explains, never in a public issue.

License

MIT. See LICENSE.

About

An AI chief of staff for product managers and founders. Runs locally in Claude or ChatGPT (Codex): one morning brief, drafts you approve.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0