MCP server

SupaNet ships an MCP server - a JSON-RPC-over-HTTP endpoint that an external Claude can connect to. This is the door that lets a tool on your own machine push agents, tools, skills, webhooks, and artifacts into your workspace, where they show up in the dashboard.

How auth works

The MCP server itself does not require a JWT, but every action is authenticated by a per-user token (mcp_tokens). You generate one in Settings → Connect Claude. Every action the external Claude takes runs as the token's owner, so it is bound by the same RLS rules as that user.

Connecting

The two common clients connect differently:

Claude Code connects directly over HTTP:

claude mcp add --transport http <name> <url>

Claude Desktop launches MCP servers as local processes, so it connects through the mcp-remote bridge configured in claude_desktop_config.json.

Settings → Connect Claude emits both snippets for you. (The desktop snippet splits the auth header via an env var to dodge a known mcp-remote bug with spaces in headers.)

What it exposes

The server exposes build tools so an outside Claude can populate your workspace:

create_agent
create_http_tool
create_skill
create_webhook
create_artifact / list_artifacts / get_artifact / list_collections / create_collection / add_to_collection / add_note (content-authoring tools - same capabilities the internal chat assistant and agents have)
create_file / list_files / get_file / delete_file / add_file_to_collection (file tools for persisting generated output - images, PDFs, text files)
list_tables / create_table / add_table_row / update_table_row (table tools for managing data tables programmatically)
the matching list_* tools

External MCP tools

Connected external MCP servers (e.g. Playwright for browser automation) are also exposed in tools/list with unified namespacing (e.g. playwright__browser_navigate). This means external AI gets both the workspace's build tools and any browser/integration tools at a single endpoint. See Playwright (browser automation) over MCP for details.

Looping system

It also exposes the looping system so you can create and run goal-directed loops from an external Claude:

create_loop — create a reusable loop (a goal + rubric + agent + feedback tool).
run_loop — start a run of an existing loop.
get_loop_run — check on a running or finished loop run.
list_loops — list the loops you can run.

In practice this means you can sit in Claude on your laptop, describe an agent or a tool, and have it appear in SupaNet ready to use. You can also kick off goal-directed runs with feedback and monitor them until they finish.

The artifact angle

create_artifact is the tool that matters for documentation. An external Claude can push HTML straight into the artifacts table and get back the same public URL the in-app editor produces - which is exactly how you can publish a page (or these docs) into SupaNet from the outside. See Teach SupaNet about itself.