# AI assistants (MCP)

You can connect the AI assistant you code with, such as Claude Code, Codex or GitHub Copilot,
to Productmind. Your assistant can then read and update your tasks, features, feature requests
and product context straight from your editor or terminal, and work on a task in your own
repository while Productmind tracks its progress. Assistants connect through the Model Context
Protocol (MCP), and you sign in once in your browser.

<Availability permission="Any role can connect an assistant, and it can do only what your role allows. Any role also sees the MCP Clients section of Integrations." />

## Key concepts

**MCP.** The Model Context Protocol, an open standard AI assistants use to reach other
services. Productmind offers its actions to assistants as MCP **tools**, such as listing tasks
or creating a feature request.

**The Productmind MCP endpoint.** The address your assistant connects to:

```text
https://api.productmind.ai/mcp
```

**Connection.** Each assistant you sign in is a connection. It acts as you, in one organization,
and stays until you disconnect it, or until it goes unused for 90 days.

## The MCP Clients section at a glance

Whatever your role, you also find the endpoint, and each assistant's setup guide, on
**Integrations**. At the top-right of any page, choose the **Integrations** icon, then
**View all**. With the **Owner** or **Admin** role, you can also choose **Integrations** in the
sidebar, under **Settings**. **MCP Clients** is the first section, or, for an Owner or Admin, the
one below **MCP Servers**.

<Screenshot id="ai-assistants-coding-agents" alt="The MCP Clients section: the MCP endpoint with a Copy button, and cards for Claude Code, Codex and GitHub Copilot" />

- **The MCP endpoint**, with **Copy**.
- **A card for each assistant**: Claude Code, Codex and GitHub Copilot. The card reads **You're
  connected** when **you** have connected that assistant, and **Not connected** until then. The
  cards show your own connections, not your teammates'.
- **Setup guide** on each card opens that assistant's own instructions for adding an MCP server.

The **MCP Servers** section above it, which the **Owner** and **Admin** roles see, is the other
direction: servers your Productmind agents use. See [Integrations](/user-guide/settings/integrations).

## Connecting an assistant

### To connect Claude Code

1. In a terminal, run:

   ```bash
   claude mcp add --transport http productmind https://api.productmind.ai/mcp
   ```

2. Start Claude Code, run `/mcp`, and choose **productmind** to sign in.
3. Your browser opens on Productmind. Sign in if you aren't already.
4. If you belong to more than one organization, choose the one the assistant should work in,
   then choose **Continue**.

Claude Code shows the server as connected, and Productmind's tools become available to it. In
your profile, under **Connected clients**, it appears with the tag **MCP**.

### To connect Codex or GitHub Copilot

1. In the assistant, add a remote MCP server with the address
   `https://api.productmind.ai/mcp`. The assistant's own documentation explains how. The
   assistant's **Setup guide**, under **MCP Clients** on **Integrations**, opens those
   instructions.
2. When the assistant asks you to sign in, your browser opens on Productmind. Sign in if you
   aren't already.
3. If you belong to more than one organization, choose the one the assistant should work in,
   then choose **Continue**.

The assistant connects. In your profile, under **Connected clients**, it appears with the tag
**MCP**.

### To connect another MCP client

Any MCP client can connect if it:
- connects to remote servers over HTTP;
- signs in with OAuth, the standard browser sign-in; and
- runs on your own computer, or is a browser extension.

Add `https://api.productmind.ai/mcp` as a server in the client, and sign in when it opens your
browser. The client registers itself with Productmind as it connects: you don't create a key or
a client ID.

:::note
Assistants that run on someone else's servers and sign in through a public web address can't
connect. Productmind only accepts a sign-in that returns to your own computer or to a browser
extension. See [API access and authentication](/user-guide/tools/api-access).
:::

### To work in another organization

A connection works in the organization you chose when you signed in. It lists only that
organization's workspaces. To use another organization, sign the assistant in again from the
assistant itself, and choose the other organization. In Claude Code, run `/mcp`, choose
**productmind**, and sign in. You don't need to remove the server and add it back.

If you're signed in to Productmind through single sign-on, you can choose only the organization
you signed in to. To connect the assistant to another one, sign out and sign in again with your
password or Google, or through that organization's own single sign-on, first.

### To disconnect an assistant

1. Choose your avatar at the bottom of the sidebar, then **Profile**.
2. Under **Connected clients**, find the assistant. Its row is labeled with the assistant's name
   and the tag **MCP**.
3. Choose **Revoke** on its row.

The assistant can't reach Productmind any more until you sign it in again. To disconnect
everything except the browser you're using, choose **Revoke all sessions**. See
[Profile and notifications](/user-guide/profile-and-notifications).

## What an assistant can do

A connected assistant can work with:

- your workspaces;
- [features](/user-guide/product/features), [feature requests](/user-guide/product/feature-requests),
  [tasks](/user-guide/tasks) and [bug reports](/user-guide/bug-reports);
- [sessions](/user-guide/sessions/overview), [releases](/user-guide/product/releases) and
  [applications](/user-guide/engineering/applications);
- [Marketing releases](/user-guide/product/marketing), which it can read but not change: what one
  announces, its sources, its public release notes and its blog posts, if your role can read
  Marketing and your organization's plan includes it;
- the [knowledge base](/user-guide/knowledge-base), and your workspace's
  [agent context](/user-guide/engineering/agents/context);
- [agentic interviews](/user-guide/feedback/agentic-interviews),
  [customer meetings](/user-guide/feedback/customer-meetings) and
  [support conversations](/user-guide/feedback/customer-support), and
  [internal meetings](/user-guide/internal-meetings);
- accounts and [contacts](/user-guide/feedback/contacts);
- comments, attachments and tags;
- [decisions](/user-guide/settings/workspace-settings/decision-records), if your organization's
  plan includes decision records;
- search across your workspace;
- the [audit log](/user-guide/settings/identity-and-access/audit-log), if your role can read it
  and your organization's plan includes it.

An assistant isn't offered the tools for a feature your organization's plan doesn't include. See
[Features that depend on your plan](/user-guide/reference/plans-and-limits#features-that-depend-on-your-plan).

To pick up a task, ask your assistant to show your tasks or to work on one. Productmind gives it
a workflow to follow: it confirms which tasks to work on, then records its plan, its progress and
its pull requests on them as it goes. Its plan, progress notes and verification, with any
screenshots it attaches to them, appear in the session's conversation under your assistant's own
name and logo (Claude Code, Codex or GitHub Copilot), so you can tell its messages from yours while
you follow the work. You can change that workflow for your workspace under
**Agents › MCP Workflow Skill**, once that tab is turned on for your organization. See
[MCP Workflow Skill](/user-guide/engineering/agents/mcp-workflow-skill).

### Application and sandbox settings

With the **Owner** or **Admin** role, your assistant can also change the settings of your
[applications](/user-guide/engineering/applications) and
[sandboxes](/user-guide/engineering/sandboxes), as you would on their pages:

- an application's install and run commands, environment variables, services and their ports,
  risk rules, and branches: **Base**, **Development**, **Production**, and the branch of each of
  your environments;
- a sandbox's applications and their startup order, its size, and its environment variables.

Your assistant lists your applications or sandboxes first to find the one to change, and each
change applies at once. It changes environment variables by name: the ones it doesn't name are
kept, and it never sees a secret's value. Attaching, switching or removing an application's
repository is done in the app.

The assistants of every other role aren't offered these tools. Neither are the PM agent, your
custom agents and automations, or coding sessions, even when an Owner or Admin started them.

## Hand a cloud session to your coding agent

A session that Productmind's agent planned or ran, in Productmind's cloud or on your
[Productmind agent](/user-guide/tools/productmind-cli), can be finished by your own coding
assistant instead. Ask your assistant to work on the session, by its task or its number (such as
**SSID-12**). It takes the session over when nothing of Productmind's agent is running there:

- A plan waiting for your approval is withdrawn, never approved, and so is a question the agent
  asked. Nothing starts in Productmind's cloud. Your assistant works from the session's
  **Engineering plan** as it reads, or drafts a new plan with you.
- The session goes back to **In Planning**, or stays in **PR Ready for Review** when it has a pull
  request. A **Failed** or **Cancelled** session is picked up again the same way.
- The session page no longer offers **Approve & build**, **Build** or **Retry**, and the session
  is listed under your name, such as **Ada Lovelace over MCP**. A line in the conversation says who
  took the session over, and what was withdrawn.

Your assistant can't take a session over while a run is in progress or waiting for a runner, while
**Agent offline** shows in the middle of a run, while a live preview is open, or once the session
is **Completed** or **Rolled Back**. It tells you why: wait for the run to end, or stop the run or
the preview, then ask again. For a few minutes after a run ends in Productmind's cloud, the
session's runner still holds it: ask again shortly. A completed session needs a new session.

A session your assistant took over stays with it: it can't be moved back to Productmind's cloud or
to a Productmind agent. The AI credits Productmind's cloud already used on the session stay on your
[usage](/user-guide/settings/billing-and-payments/usage). If your workspace changed its workflow
under **Agents › MCP Workflow Skill**, your assistant takes a session over only when that workflow
tells it to.

## Share your assistant's conversation with a session

When your assistant works a session, the session's conversation shows its plan, its progress
notes and its verification. Your assistant can also share its own conversation with the session:
your prompts, its replies, the tools it ran and what they returned, and the screenshots it kept.
The session then reads like one the Productmind agent ran, and your team can follow all of it.

These assistants can share their conversation:

- **Claude Code**, in the terminal and in VS Code or JetBrains.
- **Codex**, in the terminal and in its IDE extension.
- **GitHub Copilot**: the Copilot CLI, and Copilot Chat in agent mode in VS Code 1.137 or later.

**How it works.** Your assistant asks Productmind for an upload and runs the command it gets back
in your terminal. The command finds this conversation's history on your computer and uploads it,
and Productmind adds it to the session's conversation, in the order it happened, among the
assistant's notes. Your prompts appear as yours, and everything your assistant said and did appears
under its name and logo. Your assistant shares again at milestones, such as after its plan, after its
verification and before it opens its pull request. Each share adds only what is new. To share at
any other moment, ask your assistant to share its conversation with the session.

**You approve the upload.** It is a command your assistant runs in your terminal, so you see it
first, and your assistant asks you to approve it unless you allow its commands automatically.
Codex asks for network access to run it.

**Everyone who can read the session sees it.** The whole conversation is shared, including the
output of the commands your assistant ran and the content of the files it read. Share it only in a
session your team should see it in.

**Screenshots.** Screenshots your assistant took and kept in its conversation appear under the
tool that took them, and in the session's snippets strip above the message box. An assistant that
doesn't keep them in its conversation attaches them to its progress notes and verification
instead, and they appear under those notes and in the strip.

**Other assistants.** GitHub Copilot in JetBrains, Visual Studio or Xcode, VS Code before 1.137,
and assistants that run in the cloud keep no conversation Productmind can read. They still share
their plan, their progress notes and their screenshots.

**Not in a session the Productmind agent runs.** Its conversation is already in the session, so an
assistant can't share another one there.

## How connections behave

**An assistant acts as you.** It can do exactly what your role and your teams allow in that
organization, within what your organization's plan includes, and nothing more. If an Owner or Admin changes your role, your assistants follow
the change. See [Roles and permissions](/user-guide/reference/roles-and-permissions).

**Access lasts an hour at a time.** Each time you sign an assistant in, Productmind gives it
access for one hour, and a way to renew that access without you. The assistant renews it on its
own when the hour is up, so you don't sign in again. A connection nobody uses for 90 days
expires, and the assistant asks you to sign in again.

**A sign-in renewed from two places ends.** Each renewal replaces the one before it. Two copies
of an assistant that renew within the same minute receive the same renewal. If a replaced renewal
is used again more than a minute later, for example by a second copy of the assistant that shares
the same saved sign-in, Productmind treats the sign-in as stolen and ends it, and you sign the
assistant in again. Sign each copy of an assistant in on its own.

**Leaving ends it.** If you are disabled or leave the organization, your connections stop
working.

**What it does is recorded as you.** Records your assistant creates or changes show you as the
author, and appear in the [audit log](/user-guide/settings/identity-and-access/audit-log) under
your name.

## Permissions

| To | You need |
| --- | --- |
| Connect an assistant | A Productmind account in the organization |
| Use a tool | The permission the same action needs in the app, for example **Tasks: Create** to create a task, and, for a feature that depends on your plan, a plan that includes it |
| Work with decisions through an assistant | The permission the same action needs in the app, on a plan that includes decision records |
| Read the audit log through an assistant | **Audit log: Read**, on a plan that includes the audit log |
| Change application and sandbox settings through an assistant | The **Owner** or **Admin** role |
| See the **MCP Clients** section of **Integrations** | Any role |

## Troubleshooting

### The assistant can't find my workspace

It is connected to another organization. Sign it in again and choose the organization the
workspace belongs to. Your organizations are listed in your profile.

### The client says "redirect_uris must be a loopback address"

The client wants Productmind to return the sign-in to a public web address, which Productmind
doesn't accept. Use a client that runs on your computer, such as Claude Code, Codex or GitHub
Copilot in your editor.

### A tool fails with a permission error

Your role doesn't allow that action. Ask an Owner or Admin of your organization, or see
[Roles and permissions](/user-guide/reference/roles-and-permissions) for what each role can do.

### The assistant asks me to sign in every hour

Assistants renew their access on their own, so this usually means yours doesn't renew it.
Update the assistant to its latest version. If it still asks, remove the Productmind server from
the assistant, add it again, and sign in. If you belong to more than one organization, choose
the same one as before.

### The browser shows "Unknown client_id"

The sign-in link doesn't carry a client ID Productmind issued, for example because it was copied
by hand and cut short. Start the sign-in again from the assistant. In Claude Code, run `/mcp`,
choose **productmind**, and sign in. If you copy the sign-in link by hand, copy all of it.

### The assistant stopped working

The connection was revoked or expired, you left the organization, or its sign-in was renewed from
two places and ended. Sign the assistant in again from the assistant itself. In Claude Code, run
`/mcp`, choose **productmind**, and sign in. You don't need to remove the server and add it back.

### My assistant's conversation doesn't appear in the session

Ask your assistant to share its conversation with the session, and approve the command it runs.
If the command says it found no conversation, your assistant keeps none Productmind can read (see
[Share your assistant's conversation with a session](#share-your-assistants-conversation-with-a-session)):
ask it to attach its screenshots to its progress notes instead.

### The card says "You haven't connected this agent yet" but my teammate is connected

The cards show only your own connections.

## Related topics

- [Productmind CLI](/user-guide/tools/productmind-cli)
- [API access and authentication](/user-guide/tools/api-access)
- [MCP Workflow Skill](/user-guide/engineering/agents/mcp-workflow-skill)
- [Integrations](/user-guide/settings/integrations)
- [Profile and notifications](/user-guide/profile-and-notifications)
