MCP server

Wryte ships a Model Context Protocol server, so a coding agent (Claude Code, Cursor, or anything that speaks MCP) can research, write drafts, build animations and upload images around your posts. You write the Main version and publish.

It runs inside Wryte's Convex backend. There is no separate service to deploy, no API token to create, and no secret stored on your machine.

Endpoint

https://<your-deployment>.convex.site/mcp

Your exact URL is shown in Settings → MCP Server, with a copy button.

Connect

Choose the client you use. The setup cards on the MCP Server docs home and in Settings → MCP Server copy the exact endpoint automatically.

Claude Code

claude mcp add --transport http wryte https://<your-deployment>.convex.site/mcp

Then run /mcp, pick wryte, and choose Authenticate.

Claude Desktop

In Claude Desktop, open Settings → Connectors → Add custom connector and paste the Wryte endpoint. Claude opens the browser to finish sign-in.

Codex

codex mcp add wryte --url https://<your-deployment>.convex.site/mcp

Then authenticate the wryte server from Codex. Codex opens the browser and stores the OAuth credential in its own credential store.

Codex Desktop and Codex CLI use the same MCP configuration, so the command above also makes Wryte available in the desktop app.

Cursor

Cursor supports a one-click Add to Cursor action from the setup card. For a manual setup, add this to your global ~/.cursor/mcp.json or a project-level .cursor/mcp.json:

{
  "mcpServers": {
    "wryte": {
      "url": "https://<your-deployment>.convex.site/mcp"
    }
  }
}

After adding it, open Cursor's MCP settings and authenticate wryte when prompted.

VS Code

Use the VS Code button on the setup card, or run MCP: Add Server from the Command Palette, choose HTTP, and paste the endpoint. VS Code writes it to your mcp.json:

{
  "servers": {
    "wryte": {
      "type": "http",
      "url": "https://<your-deployment>.convex.site/mcp"
    }
  }
}

Start the server from the Extensions view or mcp.json, then sign in when VS Code asks.

ChatGPT

ChatGPT connects to remote MCP servers in developer mode, on paid plans.

1. Turn on Developer mode in ChatGPT's settings. On Business and Enterprise workspaces an admin may need to allow custom connectors first.
2. Create a new app or connector, name it Wryte, paste the endpoint, and choose OAuth as the authentication.
3. Connect, approve access in the browser, then pick Wryte from the developer mode tools in a new chat.

ChatGPT moves these menus often. If a label doesn't match, search its settings for "Developer mode". Each client opens the browser for OAuth approval; Wryte never creates or stores a copy-paste API token.

Use the URL exactly as shown, with no trailing slash. The server advertises itself under that precise spelling, and the OAuth spec makes clients verify the two match character for character.

Upgrading from the earlier server

Nothing to reinstall. The endpoint and sign-in are the same, so an existing connection keeps working and picks up the new tools the next time your client starts a session. Restart the client, or reconnect wryte, if it still lists the old tools.

What changed for agents:

• Drafts replace direct edits. wryte_drafts_create, wryte_drafts_update_content and wryte_drafts_promote are gone; use wryte_drafts_snapshot and wryte_drafts_update. Agents never touch the Main version.
• Publishing and scheduling are gone: wryte_publish_document, wryte_schedule_set and wryte_schedule_cancel. The Publish capability no longer exists, and a stored Publish grant is ignored.
• One animation tool. wryte_animations_upsert replaces create, update and replace-by-name, and runs the project's checks before saving.
• New: wryte_project_context, wryte_documents_workspace and wryte_media_upload_url.
• Stricter input. Slugs, titles, frontmatter, tags, draft labels and research are validated, and a refusal says exactly what to fix.

A saved prompt or script that calls a removed tool gets Unknown tool. Point it at the replacement above.

What an agent can do

The canonical loop this was built for:

Look at my existing posts, research this topic, file what you find, then write me a first draft.

That works out of the box: read and write are granted by default. Media upload and trash are opt-in, because uploads spend your storage provider's quota and deletion is deletion. Publishing is never available to an agent.

Design notes worth knowing

• 41 tools, deliberately. An earlier cut had 48 for posts alone. Every tool description sits in the model's context on every turn, and near-duplicate tools make models pick wrong and retry. Fewer, better-shaped tools cost less and work better.
• Nothing irreversible is reachable. No permanent delete, no project delete, no account deletion, and nothing that touches stored credentials. The worst an agent can do is trash a post whose Main version is still empty, which you can restore.
• Main is yours. Agents write draft tabs. No tool reads or writes the Main body, promotes a draft, schedules or publishes.
• Agents share your app's rules. Tools call the same functions the web app does, so every ownership check, quota and rate limit already applies.

Read next

• Authentication: how the OAuth flow works and why there's no API token
• Capabilities: the four permissions and how to change them
• Tool reference: all 41 tools with arguments
• Resources: context an agent should read before acting
• Rate limits: what's enforced, and what happens when you hit it
• Troubleshooting: every error message and what it means