These functions implement R as an MCP client, so that ellmer chats can register functionality from third-party MCP servers such as those listed here: https://github.com/modelcontextprotocol/servers.
mcp_tools() fetches tools from MCP servers configured in the mcptools
server config file and converts them to a list of
tools compatible with the $set_tools() method of ellmer::Chat objects.
Value
mcp_tools() returns a list of ellmer tools that can be passed directly
to the $set_tools() method of an ellmer::Chat object. If the file at
config doesn't exist, an error.
Configuration
mcptools uses the same .json configuration file format as Claude Desktop;
most MCP servers will define example .json to configure the server with
Claude Desktop in their README files. By default, mcptools will look to
file.path("~", ".config", "mcptools", "config.json"); you can edit that
file with file.edit(file.path("~", ".config", "mcptools", "config.json")).
The mcptools config file should be valid .json with an entry mcpServers.
That entry should contain named elements, each configuring either a local
stdio server with command and args, or a remote Streamable HTTP server
with url. Stdio MCP server processes receive an allowlisted environment
inherited from the current R process, plus any variables configured in
env. Configured env variables override inherited variables with the same
name. Servers that need additional environment variables should list them in
env.
For example, to configure mcp_tools() with GitHub's official MCP Server
https://github.com/github/github-mcp-server, you could write the following
in that file:
Connecting to remote (http) servers
For remote Streamable HTTP MCP servers, configure a server with url.
Static headers can be supplied with headers; protocol-owned headers such
as Accept, Content-Type, MCP-Session-Id, and MCP-Protocol-Version
are managed by mcptools and cannot be configured manually. Credentialed
public remote endpoints must use HTTPS. HTTP is allowed for loopback
development servers, or for explicit unsafe opt-out with allow_http.
Remote server entries support these fields:
url: the Streamable HTTP MCP endpoint.headers: named static headers. Values may use${ENVVAR}interpolation.timeout: a number of seconds for the overall HTTP request timeout.allow_http: allow credentialed non-loopback HTTP endpoints.ignore_tools: tool names or*wildcards to hide and block.oauth: OAuth settings.
OAuth settings may include authorization_server, resource, scope with
scope_mode = "override", client_info, manual_client_info,
client_metadata, redirect_uri or callback_host/callback_port/
callback_path, cache_dir, and allow_http. mcptools supports OAuth 2.1
with PKCE: it discovers the authorization server from the protected-resource
metadata advertised in a 401 challenge, registers a client dynamically when
the server supports it, and caches tokens (refreshing them automatically).
Remote HTTP requests use httr2 and curl. Proxy and corporate CA settings
should generally use the standard curl environment variables, such as
HTTPS_PROXY, NO_PROXY, SSL_CERT_FILE, and CURL_CA_BUNDLE. Stdio
server processes inherit these variables through mcptools' default
environment allowlist.
See also
This function implements R as an MCP client. To use R as an MCP server,
i.e. to provide apps like Claude Desktop or Claude Code with access to
R-based tools, see mcp_server().
Examples
# setup
config_file <- tempfile(fileext = "json")
file.create(config_file)
#> [1] TRUE
# usually, `config` would be a persistent, user-level
# configuration file for a set of MCP server
mcp_tools(config = config_file)
#> list()
# teardown
file.remove(config_file)
#> [1] TRUE
