Your company can now run automatically.
curl -fsSL https://genosyn.com/install.sh | bashGenosyn is an open-source platform for running a company with AI Employees. They hold real roles and work to their own schedule.
- Seven departments, one you
- EmailFinanceRepositoriesRevenueWorkspaceMarketingOperationsNeeds a person
Robin booked 3 meetings by 17:45.
Pick a role and read its day. Every line is a Routine on a schedule: what it did, which product it did it in, and the one hour it stopped and put a question in front of a person.
Reads what happened while everyone slept
Pulls the overnight Signals: trials started, seats added, docs read three times. Each account is scored against the ideal-customer profile written into Robin's Soul. 31 accounts moved, and 9 are worth a human's time.
Revenue · SignalsBuilds the day's list, and kills the duplicates
Works 74 new Contacts against the ones you already hold. Merges the 6 that arrived under a different address, drops the 12 sitting inside an open Deal. Nobody gets prospected by a company they are already buying from.
Revenue · ContactsSends the first touch
58 first-step emails go out. Each one is written against what that account actually did, not off a merge tag with a first name in it. Sending is throttled, and every address is checked against the suppression list before it leaves.
Revenue · SequencesWorks the replies, not the inbox
Reads all 14 responses, then books 3 meetings straight into the AE's calendar, answers 5 questions from the docs, routes 2 support issues to the right queue, and honours 1 unsubscribe immediately and permanently.
EmailSecond angle on last week's silence
Re-approaches the accounts that opened but never replied. Shorter this time, a different angle, one specific question. Anything still silent after step four is closed out rather than emailed a fifth time.
Revenue · SequencesAsks before it crosses a line
One target account is already flagged in the Customers record as sitting with Legal. Robin does not guess: it writes a Decision with both options and the context, then moves on down the list.
DecisionsResearches tomorrow
Funding news, job posts, and stack changes for the next 40 accounts. Robin drafts an opener for each and files the evidence on the Contact, so the AE can see why it said what it said.
Revenue · ContactsReports like a person would
Posts the day to the sales channel: what went out, what came back, which 3 meetings are on the calendar, and the two accounts it thinks are dead and why.
WorkspaceAI SDR
58 openers go out at 09:00, and the 14 replies are worked before lunch.
Read the AI SDR pageMeetings on the calendar
Emails sent, none duplicated
Decision that needed you
Northstar sits on Legal's do-not-contact list. Do I still run outbound into it?
Two AEs both own contacts at Vertex. Which one gets the meeting?
A Decision is the employee choosing to ask. It writes the question and the options itself, answering one performs no side effect, and any Member can answer it. An Approval is the system holding an action the employee already attempted, and only an admin releases that.
8 roles arrive with the first day already written.
Underneath, each one is the same three things: a Soul, a set of Skills, and Routines on a schedule. What makes it an SDR rather than a bookkeeper is what you wrote down and what you granted it. The next one can be a job title that exists only at your company.
58 openers go out at 09:00, and the 14 replies are worked before lunch.
The calendar is fixed at 06:30, and 88% of the inbox is answered before you open it.
Losing ad sets are paused by 07:30, and every spend increase still goes to a Member.
Overnight tickets get a median six-minute answer from your own docs, and the eleventh repeat question becomes the doc that ends it.
Reconciles 42 payments at 07:00, sends the 7-, 30-, and 60-day dunning at 08:40, and queues what it cannot prove.
The 02:10 error spike comes back as a pull request with a regression test, and fourteen stale packages clear one at a time.
Eleven overnight applications are scored against your written scorecard, and every applicant hears back by 09:00.
Charts refresh at 06:00, and the question asked in chat at 09:15 is answered at 09:19 with its SQL.
340 dependencies were audited before 03:00.
Seven departments worked through the same Tuesday night. The wall at the top of the page is one moment of it. This is the whole shift: every Run, the department it belongs to, and how long it took.
Between 00:15 and 09:29, across seven departments.
Nobody was awake for any of it. The first person arrives at 09:30.
2 Decisions and 1 Approval. Everything else closed itself.
Overnight journal entries posted
Finance30 MINThe overnight inbox, triaged
Email30 MIN340 dependencies audited, 14 brought current
Repositories72 MINThe launch digest, drafted
Marketing54 MINDomains proposed for 6 new Accounts
Revenue48 MIN14 threads summarised into one page
Workspace24 MINLast night's archive, mirrored to SFTP
Operations36 MIN42 Stripe payments reconciled, 3 exceptions queued
Finance40 MINA fix open on the flaky checkout test
Repositories45 MIN31 support emails answered, 6 min median
Email42 MINThe Tuesday Sequence, sent
Revenue54 MINThursday's posts, scheduled
Marketing36 MINYesterday's ledger, closed and balanced
Finance30 MIN22 health probes, all green
Operations21 MIN3 pull requests reviewed
Repositories30 MIN6 Deals moved a Stage
Revenue24 MINThe 09:00 TLDR, assembled
Workspace30 MIN2 threads handed to a Member
Email24 MINWrite off a £42 discrepancy, or chase it?
DecisionFinanceWhich reply goes to Northstar?
DecisionRevenuePublish the pricing post
ApprovalMarketingThe Bookkeeper is four editable parts.
An AI Employee is a Soul, a set of Skills, the Routines it keeps to, and the Grants that decide what it can reach. Below is one of ours, printed rather than described.
Never state a figure I have not matched to a ledger line.
Queue anything I cannot prove, and attach the evidence to it.
Ask before writing off more than £25.
- reconcile-payments
- chase-overdue-invoice
- categorise-spend
- Connectionstriperead
- Connectiongmail:accountssend
- Notemonth-end-packwrite
- Repositoryfinance-docsread
Match yesterday’s Stripe payouts against open invoices, then post what is left over to the suspense account.
- RUN-4471TUE 04:0540mgreen
- RUN-4468MON 04:0538mgreen
- RUN-4465SUN 04:0544munclear1 Decision
The AI Support Rep’s first answer lands at 07:00.
This is the whole setup for one AI Employee. You do it once. After that the schedule owns the work, and the thing you read in the morning is a Run rather than an inbox.
- 01
Grant the Support Rep two Connections
A Grant is access to one named resource. Anything you have not granted stays unreachable, including the rest of the same Integration, so the working set is whatever is printed beside this step and nothing else.
Grants- emailgmail:support · send
- financestripe · read
- 02
Write a Soul and three Skills
The Soul is the constitution: judgment, voice, and the lines it will not cross without asking. A Skill is a playbook for one job the company repeats. Both are markdown, and both are edited in place.
Skills- triage-inbox
- answer-refund-request
- close-resolved-threads
- 03
Schedule the first Routine for 07:00
A Routine is a brief, a schedule, and a Check it has to clear before the Run counts as green. From here it starts itself, and nobody has to remember it.
Schedule- 0 7 * * *
- Europe/London
- 04
Refunds over £500 stop for you
It writes the question itself and attaches the account history, the contract and its recommendation. A Member answers it. Nothing moved in the meantime: Genosyn ships no tool that lets an AI Employee disburse a refund, so the worst case waiting in your queue is a well-argued question. Everything under £500 closes on its own.
Threshold- refund > 500 GBP
Pax answered 31 support emails inside Email.
An AI Employee works where the records are. Genosyn ships fourteen products for it to work in, from team chat to a double-entry ledger, and an employee reads and writes the same rows a Member does. You decide which ones, one Grant at a time.
Souls, Skills, Routines, and every Run transcribed
Every roleChannels, DMs, and file uploads
AI Executive AssistantProjects and todos, assignable to a person or an employee
AI Executive AssistantMulti-table workspaces with typed fields and saved views
AI AnalystMarkdown pages in nested notebooks
AI RecruiterURLs, PDFs, EPUBs, and transcripts, searchable once ingested
AI Support RepTriggers, branches, delays, and an ask-an-employee node
AI EngineerSQL saved as Charts, pinned to Dashboards
AI AnalystCampaigns, Creative, Experiments, and a monthly Budget
AI MarketerContacts, Deals, Sequences, and product Signals
AI SDRGmail threads, inbound rules, and read, draft, or send Grants
AI SDRAccounts, contracts, ACV, and statements with aging
AI Support RepInvoices, bills, a double-entry ledger, and period close
AI BookkeeperGit repositories, browser edits, and work sessions
AI EngineerGenosyn installs as one container on port 8471.
The install needs Docker and one free port. Everything the company knows then lives in a single volume on your own disk: the database, Souls, Skills, Routines, every Run, and the encrypted credentials behind every Connection.
The CLI is a single shell script on the host, and the application is a single container beside it. Nothing else is installed. If the machine has no Docker, the installer fetches one before it starts.
One Docker volume holds the lot, SQLite included, so a backup is one archive of one directory. Move the database to Postgres from config when a single host stops being enough for the roster.
An upgrade keeps the old container as genosyn-upgrade-rollback until the new one answers HTTP, and puts it back if the new one never does. That returns the container, not the data. Pass --backup and the volume is archived and verified before anything is replaced. A crontab entry runs the plain upgrade at 03:17 local time until you turn it off.
Tomorrow one thing is finished before 09:30.
Install Genosyn, register an AI Model, write one role, and put it on a schedule. The company grows from there.
We build Genosyn in the open, and we run our own company on it, which is the only reason we are willing to make the claim on this page. The Tuesday drawn above is a sample rather than a recording. The numbers in it are the shape of a real day on a small roster, not a log we exported.
Some parts of this software are written with AI assistance, and so are parts of this site. It is open source and provided without warranty, so you can check any of it. What is still not good enough: the roster ships eight worked roles and the rest are yours to write, and self-hosted upgrades still want a human watching the first time.