Solid Growth runs an MCP server, so an agent can use the same tools as the assistant in the app. The server is live; connecting your own agent to it is not possible yet, because it accepts a signed-in browser session only.
The server
The server is at https://app.solidgrowth.com/api/mcp and speaks MCP over Streamable HTTP. Each request is handled for the person who made it, with that person's workspace and role, so a tool reaches only workspaces that person can open. Without a session it answers 401 with {"error": "Sign in to the app to use its tools."}.
Connect Claude or a local agent
Claude Code adds a remote HTTP server with:
The command saves the configuration and does not check credentials. Until agent credentials are accepted the server answers 401 and the connection shows as failed. Personal access tokens and a sign-in flow for agents are coming soon; this page will show the exact setup when they ship. An agent running in the app's own assistant panel already uses these tools.
Tools
The tools and their approval rules are declared once and shared by the assistant panel and the MCP server.
| Tool | What it does | Runs at once |
|---|---|---|
createTask | Creates a task, under a deliverable when the record is one. | Waits for approval |
createDecision | Logs a decision, open unless the person clearly decided. | Waits for approval |
draftContent | Stores a draft for the person to approve; never publishes. | Waits for approval |
A tool that waits for approval is not run over MCP. It answers that it was not run and that nothing changed; the person approves it in the assistant panel.
Look up a record
A tool that works on a record takes an optional recordUrl: the link of a record page in the app, for example https://app.solidgrowth.com/guardey/deliverables/142. The app resolves the link again with the person's own access, so a link to a record the person cannot read is refused. Looking a record up by its code (D-142, guardey/D-1) is coming soon, and so are tools that only read.
Permissions
An agent never has more access than the person it acts for. Row level security decides every read and write, a viewer's agent cannot change anything, and a write that creates or links work needs the person's approval.