* claim: #13336 * MCP server crate: an app can serve prompts, resource templates, completions and Markdown or binary resources (#13336) The shared crate (`solidcraft-mcp-server`) served only tools and JSON resources: `prompts/list` and `resources/templates/list` always answered empty, anything else was method-not-found, the `initialize` capabilities always named tools and resources, and every resource read went out as a JSON string. - `Tools` gains provided methods for resource templates, prompts (`prompts/list`, `prompts/get`) and completion (`completion/complete`); the default is none. - `initialize` names only what the app serves: resources when it has resources or templates, prompts when it has prompts, completions when it has either. - `read_resource` returns `Contents`: text or bytes, each with its MIME type; bytes go out as a base64 blob. - `prompts/get` checks the arguments against the prompt's listing before the app fills it in; completion refuses an unlisted prompt and sends the first 100 suggestions with the total. SolidCraft's replies stay byte-identical: its two resources are JSON (`Contents::json`) and it serves no prompts or templates, so `prompts/get` and `completion/complete` stay method-not-found. --------- Co-authored-by: MateuszSychXpl <151755914+MateuszSychXpl@users.noreply.github.com>
solidcraft-mcp-server
The machinery of a Model Context Protocol server: JSON-RPC 2.0 over stdio (one message per
line), the MCP lifecycle (initialize, ping, notifications/initialized), the tools,
resources, prompts and completion methods, tool definitions with their annotations, the check of
a call's arguments against the schema it was listed with, and the content blocks a tool replies
with (text, images).
It holds no app types and depends on no engine or renderer, so any app can serve its own tools
through it.
An app implements Tools: its instructions, its tool definitions as they are listed now, how a
tool runs, and the resources, resource templates and prompts it serves, if any. Server speaks the
protocol around them, and its initialize reply names only what the app serves: the tools always,
resources when it has resources or templates, prompts when it has prompts, completions when
it has either prompts or templates. A resource is read as text or as bytes, each with its MIME
type (Contents); bytes go out base64-encoded. A tools/call lists the tools, checks the
arguments against that listing (an unknown tool, a missing or unknown argument, a type or enum
mismatch is a JSON-RPC invalid-params error and the tool never runs), then calls the tool.
prompts/get checks a prompt's arguments against its listing the same way before the app fills
it in. completion/complete refuses a prompt that is not listed before it asks the app, and its
reply holds the first 100 suggestions and how many there are. Hostile input is answered, never fatal: a line longer than Info::max_line is refused
in bounded space, bad UTF-8 and bad JSON are parse errors, and serving goes on.
API
| Item | What it does |
|---|---|
Tools |
An app's tools: instructions, list, call, and optionally resources, resource_templates, read_resource, prompts, get_prompt, complete |
Contents |
What a resource read sends: Text or Blob, each with its MIME type; Contents::json(value) for JSON |
Reference |
What a completion asks about: a Prompt by name or a Resource by URI or URI template |
Server::new(info, tools) |
A server over tools; serve(input, output) until input closes, handle_line, handle, tools(), is_initialized() |
Info |
name, title, version for serverInfo, and max_line, the longest request line read |
PROTOCOL_VERSION, SUPPORTED_VERSIONS |
The protocol revisions spoken; the newest is offered for one not known |
tool(name, title, desc, schema, effect) |
A tool definition with annotations from its Effect |
Effect |
ReadOnly, Edits, Replaces, Files: the approval hints a client is given |
object(props, required), string(desc) |
Input-schema pieces: a closed object, a described string |
check_arguments(defs, name, args) |
The top-level check tools/call makes |
text_result(text, is_error) |
A tool result of one text block |
image_result(data, mime, text) |
A tool result of an image (base64-encoded here) and its text |
base64(data) |
Standard base64 with padding (RFC 4648) |