`AiMcp.ToolProvider`

Examples assume aliases Ai (@telorun/ai), AiMcp (@telorun/ai-mcp), AiOpenai (@telorun/ai-openai), and Mcp (@telorun/mcp-client).

AiMcp.ToolProvider is the bridge between MCP and agents: it wraps any Mcp.Client as an Ai.ToolProvider, so an Ai.Agent discovers and calls a whole MCP server's tools without knowing anything about MCP.

listTools() → the server's tools/list (each tool's inputSchema becomes the model-facing parameters).
callTool(name, args) → the server's tools/call; the returned content is fed back to the model. A tool error (isError) raises ERR_MCP_TOOL_ERROR, surfaced through the agent's onToolError.

The MCP content array is translated into Ai content parts block-by-block: a text block stays a text part, and an image block ({ type: "image", data, mimeType }) becomes an image part with its mimeType renamed to the contract's mediaType — so a vision MCP tool's image reaches the model as an image rather than a JSON-stringified blob. A result containing any unrecognized block kind (resource link, audio, …) is passed through untouched for the agent to serialize. A tool's structuredContent, when present, takes precedence and is fed back as-is.

This is the only module that depends on both @telorun/ai and @telorun/mcp-client — @telorun/ai stays MCP-agnostic, @telorun/mcp-client stays a pure transport.

kind: AiOpenai.OpenaiModel
metadata: { name: Gpt4o }
model: gpt-4o-mini
apiKey: "${{ secrets.openaiApiKey }}"
---
# Any Mcp.Client — stdio or Streamable HTTP.
kind: Mcp.StdioClient
metadata: { name: FilesMcp }
command: npx
args: ["-y", "@modelcontextprotocol/server-filesystem", "/data"]
clientInfo: { name: my-app, version: 1.0.0 }
---
# Bridge the server's tools into the Ai.ToolProvider contract.
kind: AiMcp.ToolProvider
metadata: { name: FileTools }
client: !ref FilesMcp
---
kind: Ai.Agent
metadata: { name: Assistant }
model: !ref Gpt4o
toolProviders:
  - provider: !ref FileTools
    prefix: "fs_"                        # → fs_read_file, fs_list_directory, …
    include: [read_file, list_directory] # optional allowlist

Manifest fields

Field	Type	Required	Purpose
`client`	ref (`Mcp.Client`)	yes	Any `Mcp.Client` (`Mcp.HttpClient`, `Mcp.StdioClient`, …).

The discovered tools are statically opaque — their set and argument schemas are known only at run time, advertised by the server. They are validated at the server boundary, not by the analyzer. Per-tool prefix/include/exclude and the dispatch loop live on the consuming Ai.Agent.

Manifest fields​

Manifest fields