Connect a coding agent to Spark via MCP

Jake Sawyer

Updated June 04, 2026 17:32

When you connect your coding agent to Productboard, the spec you wrote in Spark becomes something you can build from directly, and it always stays up to date. This guide walks product managers (PMs), product ops managers, and Productboard admins through setup and the most common workflows. You don't need an engineering background to follow it.

Note: The Productboard MCP server is in beta and included with Spark-enabled workspaces at no extra cost. Setup steps and capabilities may change while the feature is in beta.

What is MCP?

The Model Context Protocol (MCP) is an open standard that lets AI tools connect to other systems in a consistent way. You don't need to learn it. Paste one piece of configuration into your agent, sign in once, and the connection handles the rest.

Before you start

You'll get value from the MCP server fastest when two things are true:

You have specs worth pulling. The server delivers the specs you write in Spark. It can't deliver what hasn't been written yet. See Write your specs first below.
You have a coding agent. Claude Cowork, Claude Code, and Cursor are tested and documented here. Most other agents that support remote MCP connections work too.

Step 0: Write your specs

The MCP server delivers the specs you author in Spark, so the first step happens inside Productboard.

Open any initiative, feature, or subfeature and go to the Specification tab. When the field is empty, click Write spec to launch the write-spec skill. Spark reads the item's description and related context, asks a few clarifying questions, and drafts a spec into the Specification field. Review the draft and accept or reject the changes, the same way you would with any Spark edit.

A clear, current spec is what makes everything below work. The sharper your spec, the better your agent's output.

For the full walkthrough, see Write product specifications with Spark.

Step 1: Turn on MCP access in Productboard

You must be a workspace admin to control MCP access. If you are:

From the main menu, click Settings > MCP Server.
Turn on Enable MCP for this workspace.
Under Who can use it, choose which roles may connect an agent. For example, makers with admin access, makers, contributors, or viewers.

You can turn access off at any time. Doing so immediately ends the session of any connected agent.

Step 2: Connect your coding agent

Every agent connects to the same address: https://mcp.productboard.com. Sign in once through your browser using your existing Productboard login. There are no API keys to manage.

Tip: New to this? Claude Cowork is the most approachable place to start. It runs in the Claude app, so there's no terminal or code editor to set up.

Claude Cowork

Claude Cowork connects to the MCP server as a custom connector.

In the Claude app, open Customize Connectors in the left-hand navigation.
Click Add connector (the plus icon), then Add custom connector.
In the Name field, enter "productboard".
In the Server URL field, paste https://mcp.productboard.com.
Click Add.
The first time you use the connection, Claude opens your browser to sign in to Productboard. Complete the sign-in, and you're connected.

Note: On Team and Enterprise plans, a Claude workspace owner adds the custom connector once under Organization settings Connectors. Everyone else then connects from their own Customize Connectors. If you don't see the option to add a connector, ask your Claude workspace owner.

Claude Code

Tip: Use the latest version of Claude Code. Run claude update in your terminal first.

In your terminal, run: claude mcp add --transport http productboard https://mcp.productboard.com
Start Claude Code by running claude.
Run the /mcp command, choose productboard, and select Authenticate.
Your browser opens. Complete the sign-in and you're connected.

Cursor

Add the server to your MCP configuration file (mcp.json):

{
  "mcpServers": {
    "productboard": {
      "url": "https://mcp.productboard.com"
    }
  }
}

The first time you use the connection, Cursor opens your browser to sign in to Productboard. Once that's done, you're connected. Click here to learn more about MCP configuration in Cursor.

Other agents

The server uses standard remote MCP with browser sign-in. Most agents that support remote MCP connections work with the same address. Point your agent at https://mcp.productboard.com and follow its instructions for adding a remote server.

What your agent can and can't do

Your agent connects as you. It sees the same specs you can see in Productboard and nothing more. The connection respects your existing permissions, so an agent can never reach content you don't already have access to.

Once connected, your agent can:

Find specs: Search by the name of an initiative, feature, or subfeature, by who owns it, or by how it's tagged.
Read specs: Pull the full, current content of any spec you have access to.
Refine specs: Propose and apply edits to a spec, keeping Productboard up to date as decisions are made.
Ask questions: Read and post comments on a spec, so a question reaches the spec owner instead of getting lost in chat.
Report progress: Update an item's status to show where the work stands.

A couple of habits make this smoother:

Start a request with "use Productboard…" so your agent knows to use the connection. For example: "Use Productboard to show me the spec for the onboarding redesign."
Refer to specs by name, not by ID. "The notifications spec" works better than a long identifier.

Note: Your agent reads and writes real Productboard content. Comments and status changes notify your teammates, so treat them as you would any other update.

Where PMs get the most value

You don't have to wait for engineering to turn your spec into something you can click on. Here are the two workflows that save PMs the most time.

Building a prototype to test an idea early

A written spec is easy to agree with and hard to react to; a working prototype is much better. Connect your agent, point it at a spec, and turn it into a clickable prototype you can put in front of stakeholders.

Try prompts like:

"Use the spec for the onboarding redesign to build a working prototype I can click through."
"Build a rough version of the settings page from the spec, just enough to show in tomorrow's review."
"Turn the acceptance criteria in the notifications spec into a clickable wireframe."

When the prototype reveals a gap, fix it at the source. Your agent can write the change back to the spec:

"The prototype showed we never defined the empty state. Add a comment on the spec flagging this as an open question."

Hand engineering a spec that's always current

When you share a spec with your engineering team, their coding agent reads it directly from Productboard. That means they build from the version you edited today, not a copy someone pasted into a chat last week. This approach is often called spec-driven development: the spec, not a quick prompt, is the source of truth the agent builds from.

When you update a spec, the next person to pull it gets your latest thinking. That means less drift between what you wrote and what gets built, and less rework when something changes.

You can also signal when a spec is ready to pick up:

"Mark the billing spec as ready for engineering."

Tip: Your engineers may want the full setup and workflow details. See below.

Turning a spec into an update in seconds

Because your agent can read any spec you own, it's also a fast way to prepare for stakeholder updates or release notes:

"Summarize the API partners spec into one paragraph for my stakeholder update."
"Pull the billing spec and draft release notes from it."

Full technical guide for engineers

Your engineering team may want more information about the full list of tools, the data-access model, steps to revoke access, and how the server fits with spec-driven development frameworks like Spec Kit, GSD, and BMAD.

Such information can be found here.

Troubleshooting

Agent isn't using the connection? Start your request with "use productboard…" so it reaches for the right tools.
Agent replies with IDs instead of names? Ask it to use names. "Use the spec name, not the ID."
Search takes a few turns? The server finds specs by name, owner, and tags. If the first try misses, refine your wording or name the spec directly.