mcp

type Annotations ¶

type Annotations struct {
	Title           string `json:"title,omitzero"`
	ReadOnlyHint    bool   `json:"readOnlyHint"`
	DestructiveHint bool   `json:"destructiveHint"`
	IdempotentHint  bool   `json:"idempotentHint"`
	OpenWorldHint   bool   `json:"openWorldHint"`
}

Annotations are what a client is told of a tool's effects. Every hint is sent, whether true or not: Claude requires readOnlyHint and destructiveHint declared, and lets a read-only tool run without asking.

func ReadOnly(title string) *Annotations

type Caller ¶

type Caller struct {
	Account  string
	ClientId string
}

Caller is whom a tool call is for: the account whose data it reaches, and the client asking, as the request's access token names them.

func CallerOf(ctx context.Context) (*Caller, error)

type Content ¶ added in v1.98.0

type Content struct {
	Type     string `json:"type"`
	Text     string `json:"text,omitzero"`
	Data     string `json:"data,omitzero"`
	MimeType string `json:"mimeType,omitzero"`
}

Content is a block of a tool's result besides the text its structured content is given as: an image the model is shown, say. Data is base64-encoded, as the protocol carries it.

func ImageContent(data []byte, mimeType string) *Content

type Implementation ¶

type Implementation struct {
	Name    string `json:"name"`
	Version string `json:"version"`
	Title   string `json:"title,omitzero"`
}

type Reply ¶

type Reply struct {
	Status int
	Body   []byte
	Method string
	Tool   string
}

Reply is what answers a POST: a status, and a JSON body unless the status is 202. Method and Tool say what was asked, for the log.

type Result ¶ added in v1.98.0

type Result struct {
	Structured any
	Content    []*Content
}

Result is what a tool answers with to give blocks besides its structured content: Structured is what it would have answered with otherwise, and Content goes after the text of it, in order.

type Server ¶

type Server struct {
	Info         *Implementation
	Instructions string
	Tools        []*Tool
	// AllowedOrigins are the origins a request may come with. One coming with no Origin is not a
	// browser's and is let through; one with another origin is refused, which is what keeps a page
	// from reaching the server through its visitor's browser.
	AllowedOrigins []string
	// ListTtl is how long a client may keep the tool list before asking again.
	ListTtl time.Duration
}

func (s *Server) Handle(ctx context.Context, header http.Header, body []byte) *Reply

type Tool ¶

type Tool struct {
	Name         string         `json:"name"`
	Title        string         `json:"title,omitzero"`
	Description  string         `json:"description,omitzero"`
	InputSchema  map[string]any `json:"inputSchema"`
	OutputSchema map[string]any `json:"outputSchema,omitzero"`
	Annotations  *Annotations   `json:"annotations,omitzero"`

	Call func(ctx context.Context, arguments jsontext.Value) (any, error) `json:"-"`
}

Tool is a tool the server offers. Call answers with what becomes the result's structured content, a JSON object, and its text, or with a *Result to give more blocks besides; a *ToolError is a refusal, and any other error the server's.

type ToolError ¶

type ToolError struct {
	Message string
}

ToolError is a tool refusing what it was asked, told to the model as the tool's result rather than as a protocol error, so that it can try again otherwise.

func (e *ToolError) Error() string