MCP Resources as System Prompt Context with Tool-Based Fallback

Context

The McpReactAgent connects to MCP servers at runtime. MCP servers expose two primitives: tools (actions the LLM can call) and resources (read-only data identified by URI). Tools work out of the box — the LLM calls them via OpenAI function-calling. Resources do not: the LLM has no native mechanism to request a resource read.

The MCP spec is explicit about this. Resources are application-controlled — the spec states: "If you want the model to be able to request data dynamically, consider using Tools instead." In interactive clients (Claude Desktop, Cursor), a human picks which resources to attach via a UI. Our agents are autonomous — there is no human to pick.

We needed to decide how an autonomous agent consumes MCP resources.

Decision Drivers

• No LLM-native resource access
  OpenAI function-calling only supports tool calls. The LLM cannot request a resource read.
• Two types of resources with different characteristics
  Static resources have fixed URIs known at connection time. Templated resources have parameterized URIs (e.g. users://{id}/profile) whose values depend on the conversation.
• Token efficiency
  Resources that can be resolved upfront should not waste a tool slot.

Decision

MCP resources are context (like a system prompt), not actions. We handle them based on when their content can be resolved:

Static resources — URI is known, content is fixed. Fetched once at agent init, injected into the system prompt. Zero tool calls, zero overhead.

Templated resources — URI contains parameters unknown at init (can't fetch users://42/profile before knowing the user asks about user 42). Exposed as a single meta-tool read_mcp_resource so the LLM can fetch them on-demand when it knows the parameters. This is late-bound context retrieval, not an action.

Consequences

Positive

• Static resources cost nothing — they're just prompt context
• All templated resources share one tool slot regardless of count
• Follows the MCP spec's own recommendation for autonomous agents

Trade-offs

• Developers who know MCP from interactive clients may find the tool conversion surprising
• The LLM must correctly construct URIs from template patterns, which may fail for complex templates
• A server with many resource templates produces a long tool description