Comma
Weekly product sync in 15 minutes. Since last week:
- Pull request #482 merged
- ENG-219 closed
Comma gets the job done in your apps and learns how you work through personalised reinforcement learning. Run it on our service or on your own servers.
An example of Comma's Home on an instance a company runs itself, at work on what comes in from outside the company. You ask Comma to handle what came in from customers, partners and suppliers. It reads your email, your shared channels and the contracts in Drive, and checks your security docs. It lists the internal work it finished this morning, starts a Task to answer Northwind's security questionnaire from your security docs, and leaves its replies to Initech and Globex for you to send. In the Tasks rail, Legal reviews Globex's contract changes against your standard terms, Security prepares for the audit, and Finance pays this month's supplier invoices. This week it answered 64 customer emails and 31 partner questions, reviewed 9 contracts and updated 14 supplier records.
Train your first private AI model with RLaaS.
Comma for Teams is your AI partner. It works in your apps, takes on tasks, and gets the job done.
Grows with the teams and people that use it.
An example of RLaaS at work: your model, at version 3, labels a refund request in Linear as a bug. Priya corrects the label to Billing. The model learns from your team's corrections, and version 4, which beats version 3 on your test cases, labels the next refund request Billing itself.
SOTA-level capability for your business: more kinds of tasks, a better understanding of how you work, and safer execution.
RLaaS trains your first private AI model through personalised reinforcement learning, using your team's work and feedback.
Grows with the workflows it learns and where it runs.
Comma follows the channels you choose. It answers questions, shares notes and drafts the replies that you approve.
Comma triages and assigns new issues, and it closes an issue when its fix merges. A decision that is yours waits for you.
In Gmail and Outlook, Comma drafts replies, files invoices and unsubscribes from newsletters. Set it to draft only, and you send each reply yourself.
Comma keeps plans, notes and checklists in Notion current, with a link to each source. Changes that need review wait for you.
It joins Google Meet calls on the calendars you choose, keeps the transcript, and turns the call into notes and next steps.
Comma posts a brief to the team's channel: what changed since the last meeting, with links.
Comma joins as a named participant and keeps the transcript.
The summary, action items, transcript and recording arrive in one thread. Ask Comma there to file the action items in Linear.
A large model thinks a problem through. A small model makes what your team has already worked out cheap and steady.
When your team does a kind of work the same way every time, the small model learns it and takes over that workflow, as your contract sets. All other work stays with the large model.
We use the record of the agents' tasks: what they did, what your team approved, changed or turned down, and the outcome.
We turn those decisions into test cases and a score for good work, and keep both current each month.
We train a small model against those reward signals. We release a new version only when it beats the version in use on your test cases.
We deploy the new version on the workflows your contract names. All other work stays with the large model.
We host Comma for Teams: the dashboard, chat connections and task history run on our service. The agents do their work on your own computers or on a cloud computer for your team.
People sign in with your company's single sign-on through OpenID Connect, Feishu or Lark. Admins set roles, choose the models and provider keys the team uses, and export the audit log.
Apps connect with scoped permissions that you can revoke. Agents never see raw credentials.
SecurityRLaaS trains your model on your data, and only under the RLaaS contract. We do not use your data to train general-purpose models, ours or anyone else's.
Data Use PolicyComma reads mail, issues, documents and calls to do its work. Self-hosted, it reads them on servers you manage.
Tasks, memory, files and transcripts stay in your own PostgreSQL, ClickHouse and object storage. Model calls go to the provider you choose, with your key.
Self-host guideOne Docker Compose file starts the web app, the Admin console, the server and its data stores on a machine you choose.
Set your provider's endpoint, model and key, or let each person add a private model. Calls bill to your provider account, not to Comma credits.
Sign-up stays off until you turn it on. In the Admin console, the owner creates accounts and chooses which models people can select.
For each person who works with the agents
$80 / seat / month
Billed quarterly, from 100 seats
Renews each quarter until you cancel
Buy seatsTrain your first private AI model with RLaaS.
Custom pricing
Compute quota, update scope and SLA set in your contract
Contact salesRLaaS is continuous learning, not a one-time delivery. Each month we keep your data and evals current, release a new version when it beats the version in use, and track quality and cost per task.
Comma for Teams is billed per seat each quarter, through Stripe. Model subscriptions and API usage are paid separately. RLaaS is billed per company, not per seat. They are two subscriptions, not one bundle.
Initial training and integration can be quoted on their own, apart from the subscription.
Linear, GitHub, Gmail, Google Calendar, Google Drive, Notion, Outlook and Slack, each through its own sign-in. Some need an app registration with the provider in your name, and we help you set those up. Plugins add MCP servers and Skills for other tools.
Google Meet calls on the calendars you choose, or a call you send it a link to. The meeting runtime is not part of the Compose stack, and we set it up with you.
Much of a team's work is settled: the same kind of decision, made the same way, many times a day. A small model trained on your decisions does that work faster, for less, and more consistently. The large model stays for what is new.
You do not need to. RLaaS includes the compute for training and evals, up to the quota in your contract. Compute beyond the quota is billed separately.
One seat is one person who works with the agents, whatever their role. Engineers, product managers, support, operations and sales count the same. Buy at least 100 seats. The plan renews each quarter at the same seat count until an admin cancels it. RLaaS is priced per company, not per seat.
Comma for Teams costs $80 per seat a month, paid each quarter through Stripe, from 100 seats. RLaaS is quoted for each company, by the workflows it covers; your contract sets its compute quota, update scope and SLA. Initial training and integration can be quoted on their own.
The Comma web app and Admin console, the Comma server with its agents, PostgreSQL, Redis, ClickHouse, and MinIO for stored files. One Docker Compose file starts them on one machine, and sign-in codes go out through your mail server.
The model provider you set for the instance: its endpoint, protocol, model ID and key. People can also add a private model in Settings. Calls bill to your provider account, and a self-hosted instance needs no Comma credits.
Not unless you ask. We do not use your files, conversations or prompts to train our own or third-party general-purpose AI models. Training on your data happens only in a service you ask for, such as RLaaS, under its own contract. Self-hosted, your data stays in your stores, and model calls go to the provider you set.