Give your AI coding agent your game's design context (MCP)

The problem: the agent codes fast, but it doesn't know your game

When you ask a coding agent for a prototype, it only knows what you paste into the conversation. But the game's intent, the references, the ideas you dropped and the decision you made after brainstorming live somewhere else. So you re-paste the same context every session, or the agent invents a mechanic that doesn't fit your game. Impstash keeps that context in one place (reference board, ideas, project direction, tasks), and its MCP server makes it available to the agent.

1. Prepare the context in Impstash

The agent is only as good as what the project contains. Gather your references on the board, brainstorm, mark the ideas you keep as kept, then fill in Settings → Project direction: pitch, genre, pillars, tone, audience, constraints and notes. If the direction is empty, the server tells the agent, which then has to infer the game from the board notes, so it is better to write it. From a kept idea, the create-task button uses its title and text as a starting point.

2. What the agent can read (all plans)

The MCP server exposes read tools. get_project_context is the one to call first: it returns the project direction (pitch, genre, pillars, tone, audience, constraints), kept ideas, board notes, the names and URLs of reference images, links (Steam, YouTube…) and open tasks. The others: list_projects, get_project_overview, list_ideas, list_tasks, list_my_tasks, get_task_context, get_feedback, get_activity, read_board and view_board_image, which lets the agent look at an image from the project. Reading is included in every plan, Discovery included.

3. What the agent can write (Creator and Studio plans)

Writing is reserved for paid plans. For tasks: create_task, update_task, claim_task (take a lease before working), release_task and submit_deliverable (send a note and links to commits or files for human review). With a full-write token, the agent can also use add_idea, update_idea_status, vote_idea, add_board_note and add_board_link. On Discovery, the agent reads but changes nothing. It cannot move a task to "Done" unless you allow it in the project settings: it submits the task for review instead.

4. Create an access token

In Impstash, open Settings → Agents & MCP. Name the token (for example "Claude Code – desktop PC"), choose its scope (read-only, read + modify tasks, or full write) and the project it covers (this project or all your projects). The token starts with imp_ and is shown only once: copy it right away. You can revoke it from the same screen. Start with a token limited to one project and to the scope the agent really needs.

5. Connect Claude Code or another client

The MCP server answers over HTTP at https://impstash.com/mcp and expects an Authorization: Bearer header followed by your token. In Claude Code, one command is enough. For Cursor, Codex or any client that accepts a remote HTTP MCP server with headers, add the second block to the client's MCP configuration (check its documentation for where the file lives). Then check the connection by asking the agent to call list_projects.

claude mcp add --transport http impstash https://impstash.com/mcp --header "Authorization: Bearer imp_YOUR_TOKEN"

{
  "mcpServers": {
    "impstash": {
      "type": "http",
      "url": "https://impstash.com/mcp",
      "headers": {
        "Authorization": "Bearer imp_YOUR_TOKEN"
      }
    }
  }
}

6. The recommended workflow

The server hands these instructions to the agent itself. It first reads get_project_context, then list_my_tasks to see what is assigned to you. It takes a lease with claim_task, reads get_task_context (description, linked idea, previous versions and review feedback), works in your repository with its own tools, then sends submit_deliverable with a short note (what changed, how to test) and links. You accept or ask for a redo; in that case the agent reads get_feedback, claims the task again and fixes it.

7. Staying in control

The agent acts for the owner of the token, and its actions are attributed to the agent in the project activity. A token can be limited to one project, stay read-only, or be revoked at any time. The project setting "Agents can create and edit" cuts off all writing. Requests are limited to 180 per minute per account. Impstash does not run your code or modify your repository: the agent does that with its own tools, under your supervision.

A prompt to adapt to your game

Use the Impstash MCP. Call list_projects, then get_project_context for Dicebound and read the direction, the kept ideas and the constraints. Then call list_my_tasks, take the top task, claim it with claim_task and read get_task_context. In this repository, implement only what the task asks, respecting the project's pillars and constraints. If the context contradicts the task or information is missing, stop and ask me. When you are done, call submit_deliverable with a note (what changed, how to test) and the commit link. Do not move the task to "Done" yourself. If writing is refused, tell me instead of working around it.

Replace the references and constraints with your own. The AI’s answers are leads to review; the choice remains yours.

References and communities to keep going

• Model Context Protocol — The open protocol agents use to connect to external tools and data.