MCP Tool Agent

This folder contains a minimal agent example that calls a single, specific MCP tool directly using ram_client.invoke_tool(...), instead of delegating tool selection to the LangGraph agent as the default tool calling agent does. This is useful when you always want a deterministic call to one known tool rather than letting an LLM decide whether/when to call it.

Input to This Agent Experiment

Retrieval Settings - System Prompt

No system prompt is required. This template never calls an LLM — it parses the user’s message directly and invokes the tool.

Tools

Configure the tools you want available from any MCP Tool Server — any tool registered on the server(s) can be called by name.

You still need to add the MCP Tool Server to this agent experiment in RAM; only tools from servers attached to the experiment are discoverable via ram_client.get_langchain_tools() and callable by prompting.

Collections

No collections are required for this example.

Environment Variables (optional)

• TOOL_TIMEOUT: The timeout, in seconds, passed to invoke_tool (defaults to 30.0).

Set this on the Experiment’s Environment Variables page, not on the Code Template — the Code Template only holds run.py, while environment variable values are configured per Experiment. Republishing the template does not set or carry over this value.

No other environment variables are required. The tool server ID is derived automatically from the tool name you provide, using ram_client.get_langchain_tools().

Agent-specific Configuration

No code changes are required; set the TOOL_TIMEOUT environment variable on the Experiment (not the Code Template) to override the default.

Runtime Behavior

1. On the first message of a session, the template replies with instructions asking the user for a tool name and its arguments.
2. On subsequent messages, the template recognizes a few commands:
   • list tools (or any message mentioning “tool” plus “list”/”available”/”all”/”what”): returns a bulleted list of all available tool names.
   • args <tool_name> (or any message mentioning “arg” plus a tool name): calls that tool with no arguments and returns the result (or the server’s error), since that’s the most reliable way to reveal a tool’s required parameters — the client-side tool schema only exposes a generic args/kwargs stub, not real per-parameter info. If multiple tool names appear to match, the longest (most specific) one wins, e.g. a tool named getTeamsTeamId over getTeams. The response also inspects the result text for error indicators, since some MCP servers report failures as text in the result instead of raising.
   • <tool_name>, <arg1>=<value1>, <arg2>=<value2>, ...: invokes the named tool with the given args.
   • Anything else (including a message on the first turn, or one that doesn’t match a known tool) returns the help text.
3. For a tool call, it looks up the matching tool’s tool_server_id from ram_client.get_langchain_tools() and calls ram_client.invoke_tool(tool_server_id=tool_server_id, tool_name=tool_name, args=args, timeout=TOOL_TIMEOUT).
4. The template extracts the text content from the tool result, pretty-printing it if it’s JSON, and returns it.

Automation Hook (Optional)

There are no automations for this agent.