Fusion Documentation & Source MCP Server

Why connect it?

Once this MCP server is wired up, Claude or ChatGPT can answer almost any Fusion-related question in under 30 seconds. It makes the assistant aware of Fusion's concepts — compute services, invalidation, etc. — and lets it dig into the source that implements them. In our experience this works better and faster than instructing the assistant via CLAUDE.md (or a system prompt) to read Fusion's source and documentation from a nearby folder.

The Fusion documentation site exposes a public, read-only Model Context Protocol server. It gives AI coding assistants focused access to the Fusion documentation (the same Markdown and anchors used by this website) and to the Fusion source code (the .cs / .razor files under src, samples, and tests).

Endpoint: https://fusion.actuallab.net/mcp

No authentication or API key is required. All tools are read-only and return Markdown or plain text.

Documentation tools

Tool	Inputs	Result
intro	None	A compact explanation of Fusion's mental model and documentation coverage.
search	query, optional limit (default 10, maximum 20)	Ranked documentation titles, URLs, and exact anchors.
get	anchor	The Markdown for one anchor. Small sections are returned in full (including sub-headings); large sections return the immediate text plus links to their sub-headings.
search_expanded	query, optional limit (default 5, maximum 10)	Ranked matches expanded through the next heading at the same or a higher level.

Source-code tools

Two entry points, mirroring how you would navigate a checkout: locate a file (or a declaration) first, then read it.

Tool	Inputs	Result
source_index	pattern, optional limit	Files whose path or top-level type names match a regex — use it to find the file(s) you need.
symbol_search	pattern, optional limit	Declarations (types, methods, properties, fields — any accessibility) matching a regex, each with its file and line range to fetch.
source_search	query, optional context, fixedStrings, ignoreCase	ripgrep over the source, bounded to 64 KB of output and ~1 s.
source_read	file, optional startLine, endLine	A whole source file (capped at 64 KB) or a specific line range of it.

For heavy source navigation, cloning github.com/ActualLab/Fusion and using your own tools is usually faster; this MCP is best for conceptual documentation and quick source lookups.

The introduction is also available as a normal page at ActualLab.Fusion in Brief, but it is intentionally not included in the documentation sidebar.

Claude Code

Add the remote server for the current project using Claude Code's default local scope:

claude mcp add --transport http fusion-docs https://fusion.actuallab.net/mcp

Or add it once at user scope to make it available in every project:

claude mcp add --transport http --scope user fusion-docs https://fusion.actuallab.net/mcp

Local configuration takes precedence if a server with the same name also exists at user scope.

Verify the configuration:

claude mcp get fusion-docs

Start Claude Code and run /mcp to inspect the connection. You can then ask, for example:

Use the Fusion documentation MCP server to explain computed-state update delays and show the relevant API types.

To share the server through a repository instead, add this entry to .mcp.json:

{
  "mcpServers": {
    "fusion-docs": {
      "type": "http",
      "url": "https://fusion.actuallab.net/mcp"
    }
  }
}

Codex

Codex CLI stores MCP server definitions in ~/.codex/config.toml. On recent versions you can add the remote server with a single command:

codex mcp add fusion-docs --transport http https://fusion.actuallab.net/mcp

Or add the entry to ~/.codex/config.toml yourself:

[mcp_servers.fusion-docs]
url = "https://fusion.actuallab.net/mcp"

If your Codex build predates native streamable-HTTP support, bridge the remote endpoint through mcp-remote as a stdio server instead:

[mcp_servers.fusion-docs]
command = "npx"
args = ["-y", "mcp-remote", "https://fusion.actuallab.net/mcp"]

List the configured servers to verify:

codex mcp list

Start Codex and ask, for example:

Use the Fusion documentation MCP server to explain computed-state update delays and show the relevant API types.

ChatGPT

ChatGPT currently connects to custom remote MCP servers through developer mode on supported Business, Enterprise, and Edu workspaces:

1. Have a workspace admin enable developer mode under Workspace Settings → Permissions & Roles → Connected Data.
2. Open Settings → Apps → Advanced Settings and enable developer mode for your account.
3. Select Apps → Create and enter https://fusion.actuallab.net/mcp as the MCP endpoint.
4. Select no authentication, then choose Scan Tools.
5. Create the app and enable the resulting draft app in a new chat.

The exact labels can change while custom MCP support is in beta. ChatGPT must connect to the deployed HTTPS endpoint; it cannot connect directly to a server running only on your development machine.

Suggested Workflow

For an unfamiliar Fusion task, call intro once, use search to locate likely sections, and then get or search_expanded for the most relevant anchors. To work with the implementation, use source_index or symbol_search to locate files or declarations, source_read to read the exact region, and source_search to grep across the source. This keeps the model context focused while retaining direct links to the complete website documentation and source.