Skip to main content

For software developers

Coding agents in your codebase, on your team's terms.

Productmind runs every coding session in a sandbox of its own, under your team's instructions, models and branch rules, and brings the work back as pull requests your reviewers can read in order.

Ready for agents

Set up each repository once.

Each repository you add becomes an application, with its install and run commands, its services and ports and the names of its environment variables filled in. The same scan checks how ready it is for coding agents, and documents it.

  • Installs that must work.

    Turn on Require dependencies to install, and a plan, a build or a scan stops before the AI starts when the install fails, takes over 10 minutes or can't run, and says why.

  • Private packages too.

    Store a registry token as a secret variable and use it in the install command. Sessions, live previews and scans all install with it.

  • What the folders don't say.

    Scan guidance tells the scans which folder is the leading implementation, or which code is being replaced, over what the folder layout suggests.

  • A readiness report.

    Each repository is rated for coding agents out of 100, with a finding and a recommendation for each area, from CLAUDE.md and AGENTS.md to agent skills and documentation freshness.

Read moreFeature discoveryApplications in the user guide

Meet Software Inc, a fictional company that sells invoicing software. Every name in these pictures is sample data.

Where it runs

A sandbox your team defines.
A runner no one else uses.

A sandbox says which applications a session clones and starts, in which order, on which image, size and variables. In Productmind's cloud, each session runs on a runner that serves only that session, never another session or another organization.

  • Your stack, started in order.

    Attach the applications the work needs. A session clones each one and starts them in one container, in the order you set, each on its own ports.

  • Your image and variables.

    Use Productmind's standard image or your own. Variables and secrets reach every session in the sandbox, and an application's own value wins over the sandbox's.

  • Three sizes.

    Small, Medium or Large: 2, 4 or 8 vCPU with 8, 16 or 32 GB of memory. Your plan sets the hours of each it includes.

Read moreSandboxes in the user guide

Sample data

Your agent

Keep Claude Code. Run more of it.

Install the Productmind CLI and start pm agent, and your machine becomes a place to build sessions. Each one runs the coding agent you already use, Claude Code or Codex, on your own account, and opens its pull requests just as a cloud session does.

  • Your account, not our credits.

    A session on your machine uses none of your AI credits, for a fixed fee per session, and moving it between your machine and the cloud never charges it twice.

  • Three at a time, or more.

    Your agent works on up to three sessions at once, or as many as 16 if you set it, and never on the same session twice.

  • Your name on the work.

    On your machine, commits carry your own Git name and email. Connect your own GitHub, GitLab or Bitbucket account, and the pull requests open as you.

  • Or in our cloud.

    Choose Build on the cloud, and the session runs on a runner that serves only that session. A Live session goes to your machine first while your agent is online.

Read moreProdutyping and coding sessionsThe Productmind CLI in the user guide

Sample data

What it follows

Your conventions, written once.
Read by every session.

Every session plans and builds by your workspace's instructions, can load the skills your team wrote, and runs on the models you choose, on top of the specifications, business rules and documents of the work it is given.

Model classes and your own model key depend on your plan.

Read moreSpecs and documentationAgents in the user guide

Sample data

What it may touch

Your branch rules hold.
Some repositories are only read.

Before the agent builds, Productmind checks that it can push to each of the session's repositories. When an account can only read one, or a branch rule would refuse the push, the build waits and says why until you fix it and choose Check again.

Sample data

  • Read-only repositories.

    Make a repository read-only for every session, in its application, or for one session. The agent reads its code for context; nothing is committed, pushed or opened there.

  • Workflow files wait for a person.

    Your GitHub Actions workflows run with your repository's secrets, so the agent can't push changes to them. Someone on your team approves the exact content first.

  • In the creator's name.

    A session's pull requests open as the person who created it, when their own account can push, and its commits link to that account. Your workspace can have that account push the code, too.

  • Deploy branches left alone.

    A branch mapped to an environment records a deployment when work merges into it, and coding sessions don't work on it.

Read moreSecurity and governanceWho pushes a session's code

Checked before review

It runs the app before you read the diff.

A session's sandbox clones and starts your applications in the order you set. When your app has a web front end, the agent opens it in a browser inside the sandbox and takes a screenshot of each page or state its change affects.

  • The screenshots stay.

    They're kept in the session's conversation, in one strip above the message box, so you see what changed before you open the code.

  • Try it yourself.

    Start a live preview of the app the session is changing, inside the session's page, and ask the agent for a change while you watch.

  • Nothing left running.

    A database or a file watcher the agent left running is stopped a minute after it finishes, and the session's timeline says what was stopped.

Read moreProdutyping and coding sessions

Sample data

How it comes back

Know which files to read first.

Each changed file is flagged by its application's risk rules. By default, data migrations, API contracts, authentication and CI files are high risk, and tests, documentation and lockfiles are never flagged. Change the rules for each application, and type a path to see how it would be flagged.

  • Its own failing checks, fixed.

    When checks fail, the agent fixes the failures its own changes caused, at most twice a session and once a commit. When the base branch fails the same checks, it doesn't try.

  • Your CI runs once.

    Each finished step is pushed to the session's branch, and an open pull request turns draft while the agent works on it. A CI that skips drafts runs once, at the end.

  • Conflicts are a turn.

    When the base branch conflicts, Resolve conflicts automatically has the agent merge it in and push a merge commit. No commits are rewritten.

Read moreRisk rulesPull request checks

Sample data

Review

Review it like a teammate's pull request.

Each repository the session changed gets one pull request in GitHub, GitLab or Bitbucket, and you review it in Productmind, with its checks beside it.

Sample data

  • Comments become commits.

    Comment on the lines you want changed and choose Apply code revisions. The agent changes the code on the same branch and pushes to the same pull request.

  • You merge.

    The documentation and specifications the change made out of date come as pull requests of their own. Only a person merges, and merging every pull request completes the session and its tasks.

Read morePull requests and reviews

When a run goes wrong

Errors you can act on.
Work that isn't lost.

Every error in a run says what happened, why and what to do, and keeps what the failing program printed. Owners and Admins see the workspace's errors together: how often, at which stage, in which repository and application.

  • A lost runner is replaced.

    If its runner stops during a run, Productmind starts a new one, once, and the run continues there.

  • Nothing done twice.

    A build that moves to another machine picks up from the steps already pushed to its branch, instead of doing them again.

  • An offline machine waits.

    When a developer's machine goes offline mid-run, the session continues once their Productmind agent is back, or you move it to another agent or to the cloud.

Read moreCoding errorsReading a session's errors

Sample data

Just build it.

An idea or a backlog. Either way, you're not building alone.

Get started