Meet Cloptima Duet, the AI Teammate That Follows Your Rules
Tag @Duet in Slack. It works on your code in its own sandbox and opens draft pull requests, and every model call runs through your org's policies and guardrails.
In this post
An AI teammate under your rules
Teams want an agent that does real work: fix the failing test, update the dependency, write the migration.
Security leads want to know what that agent can reach before anyone tags it. Duet answers both with the rules your org already set for its AI.
How it works in Slack
Anyone on your team tags @Duet.
It does the work in its own sandbox and replies in the same thread.
Picture this
Jake tags @Duet in #payments: the checkout test fails after the coupon change. Duet clones the repository in its own sandbox, finds the bug, runs the tests, and opens a draft pull request from its own branch. A reviewer asks for one more test and mentions @Duet on the pull request. Duet adds it and replies in the same Slack thread.
- Tag @Duet in a thread. It reads the thread and replies there, with files.
- It clones your repos, runs the tests, and opens pull requests, always as drafts.
- A failed CI check, or a reviewer who mentions @Duet on its pull request, starts a new task in the same thread.
- It picks a thread up days later where it left off, and keeps your team's notes in Slack canvases.
- Save a thread as a team skill with one click, and Duet reuses it.
Four rules outside the agent
Its limits are fixed rules outside the agent: the same answer every time, whatever a prompt says.
| Rule | What it means |
|---|---|
| Only the repos you grant | Duet pushes only to its own branches, in the repos you grant, and opens its pull requests as drafts. It can't merge, and it can't change your CI workflow files. |
| Never your keys | GitHub tokens and model keys are added on the way out, never inside its sandbox. The sandbox holds only a short-lived key that reaches Cloptima and nothing else. |
| Your guardrails on every call | Secrets and personal data are masked before a model sees them, and prompt attacks in what Duet reads are caught, with guardrails from Azure, Google, AWS, your own webhook, or your own rules. |
| Hard limits, set by your org | A push to main is refused before any access is granted. The models Duet may use, its spend limits, and an off switch are yours to set. |
Your admins set it once
Duet's model calls run through Cloptima, like the rest of your AI traffic, so your admins see and control it in one place.
- Choose the model: Claude, GPT, or Gemini, with your own API keys or Cloptima credits.
- Set spend limits: daily and monthly hard limits, chosen by your admins.
- See every task: each step is logged with its model, tools, and cost.
- Turn it off: the next request is refused at once.
Coming soon
Duet works on code today. Here is what comes next.
- Duet will investigate cloud costs, logs, resources, and Kubernetes, read-only first, under a role you grant.
- Duet will work in Jira, Linear, Notion, Google Drive, and your own MCP servers.
- Duet will run inside your own cloud, so its sandbox, credentials, and model calls stay in your network.
Early-access teams get it first.
Get early access
Duet is in early access on the AI Growth, Platform Growth, and Enterprise plans.
Ask for early access, and we set it up with you. The launch film is at the top of this post, and the Duet page has three short films that follow Duet's work hop by hop.