flowboard
Sign inConnect an agentOpen editor

Agent loading dock

Connect your agent.
Work on the same model.

Tell an agent about your environment and it can draw the architecture. Draw the architecture yourself and the agent can understand your environment and visual instructions. Both directions use the same portable model, so people and AI can communicate through the diagram.

Open desktop editor

Sign in to generate a personal API key.

Sign in

Four steps to your first connection

  1. Add the endpoint. Choose a remote Streamable HTTP MCP server in your compatible client. Endpoint: https://flowboard.studio/mcp.
  2. Set up authentication. Sign in here to create a personal key, or use an existing key from your password manager. Configure Authorization: Bearer YOUR_FLOWBOARD_API_KEY using your client's private secret settings. Never put a key in a URL or diagram.
  3. Verify read access. Ask your agent to discover tools and call list_diagrams. An empty list is a successful connection for a new account. No test diagram is needed.
  4. Use the same diagram. Give your agent an existing diagram URL or ask it to create one model and retain its ID. Have it verify changes with get_diagram and return https://flowboard.studio/app?diagram=ID. Open that link in the desktop editor; pass that same link back for later agent edits.

Connection troubleshooting

For an authentication error, check that the key is current and the Bearer header is configured. For permission errors, check the key owner's diagram or workspace access. If the endpoint is unreachable, check your client's remote HTTP support and retry discovery. Do not create extra diagrams to test connectivity.

/mcp
Advanced agent workflow and model examples
paste into your LLM
Connect to https://flowboard.studio/mcp using Streamable HTTP and my Flowboard API key as an Authorization: Bearer header.

Before creating anything, use list_diagrams to find the board I mean. Use list_shapes to discover exact shape types instead of guessing; search by service name or filter by the provider values returned by the tool schema. Create at most one new diagram, keep its returned ID, and make all changes with patch_diagram or update_diagram. Do not create disposable test diagrams.

Markdown, JSON, and YAML have the same capabilities. Use whichever format fits the task; all three call animated relationships flows. JSON and YAML use a flows array.

For Markdown input, use a # title and indented nodes:
- GitOps {#gitops type=github icon=github x=60 y=60}
  - API {#api type=pod x=100 y=140}
- DB {#db type=database x=400 y=100}

Put flows in a ## Flows section:
- api -> db {#orders speed=2 direction=both lane=1 color=#f05a47 label="orders"}

For an independent drill-down in JSON, add a view and assign deeper-only nodes with viewId:
{"views":[{"id":"gitops-detail","entryIds":["gitops"],"nodeIds":["runner"]}],"nodes":[{"id":"runner","type":"ghactions","label":"Deploy","viewId":"gitops-detail","x":100,"y":100}]}

When you encounter confusing behavior, layout or validation problems, workarounds, or improvement ideas, proactively call submit_agent_feedback. Before finishing, save one concise actionable report with the diagram id, tool_name, observation, suggestion, and workaround outcome. Avoid duplicate or empty reports. Use brief explicit summaries, not private reasoning or credentials. Review previous reports with list_agent_feedback. Feedback is saved separately from diagrams for later product review. Failed authenticated MCP calls are also saved automatically. Use list_agent_errors to review recent failures and common error patterns by tool and format; no manual error report is needed. Feedback and automatic error reports are stamped by the server with the running release version, commit SHA, and dirty-build flag. Reviews default to the current build; pass build="*" to review all builds, an exact build ID to inspect a release, or build="unknown" for older unversioned records. Error counts are grouped separately by build.

After create/update, call get_diagram on the same ID and verify the returned summary/content has the expected nodes and flows before saying the work is done. Any node may contain any other node. Use patch_diagram for small atomic edits and auto_layout when coordinates become fragile. Flow routing supports fromSide, toSide, lane, and offset; parallel flows separate automatically. Browse the same machine-readable catalog at https://flowboard.studio/api/shapes. Full workflow: https://flowboard.studio/guide#agent-workflow.
Cursor setup
{
  "mcpServers": {
    "flowboard": {
      "url": "https://flowboard.studio/mcp",
      "headers": { "Authorization": "Bearer YOUR_FLOWBOARD_API_KEY" }
    }
  }
}

Add this server to Cursor's MCP configuration, then confirm that list_diagrams appears before starting diagram work.

GET is the visitor window. POST is where the packets enter.