Anubis.Server behaviour (anubis_mcp v2.2.0)

Copy Markdown

Build MCP servers that extend language model capabilities.

MCP servers are specialized processes that provide three core primitives to AI assistants: Resources (contextual data like files or schemas), Tools (actions the model can invoke), and Prompts (user-selectable templates). They operate in a secure, isolated architecture where clients maintain 1:1 connections with servers, enabling composable functionality while maintaining strict security boundaries.

Quick Start

Create a server in three steps:

defmodule MyServer do
  use Anubis.Server,
    name: "my-server",
    version: "1.0.0",
    capabilities: [:tools]

  component MyServer.Calculator
end

defmodule MyServer.Calculator do
  @moduledoc "Add two numbers"

  use Anubis.Server.Component, type: :tool

  schema do
    field :a, :number, required: true
    field :b, :number, required: true
  end

  def execute(%{a: a, b: b}, _frame) do
    {:ok, a + b}
  end
end

# In your supervision tree
children = [{MyServer, transport: :stdio}]
Supervisor.start_link(children, strategy: :one_for_one)

Your server is now a living process that AI assistants can connect to, discover available tools, and execute calculations through a secure protocol boundary.

Capabilities

Declare what your server can do:

  • :tools - Execute functions with structured inputs and outputs
  • :resources - Provide data that models can read (files, APIs, databases)
  • :prompts - Offer reusable templates for common interactions
  • :logging - Allow clients to configure verbosity levels

Configure capabilities with options:

use Anubis.Server,
  capabilities: [
    :tools,
    {:resources, subscribe?: true},      # Enable resource update subscriptions
    {:prompts, list_changed?: true}      # Notify when prompts change
  ]

Extensions are declared by identifier with their settings map. They are advertised from protocol revision 2026-07-28, which added extensions to the server capabilities:

use Anubis.Server,
  capabilities: [:tools, extensions: %{"io.modelcontextprotocol/ui" => %{}}]

Components

Register tools, resources, and prompts as components:

component MyServer.FileReader           # Auto-named as "file_reader"
component MyServer.ApiClient, name: "api"   # Custom name

Components are modules that implement specific behaviors and are automatically discovered by clients through the protocol.

Server Lifecycle

Your server follows a predictable lifecycle with callbacks you can hook into:

  1. init/2 - Set up initial state when the server starts
  2. handle_request/2 - Process MCP protocol requests from clients
  3. handle_notification/2 - React to one-way client messages
  4. handle_info/2 - Bridge external events into MCP notifications

Most protocol handling is automatic - you typically only implement init/2 for setup and occasionally override other callbacks for custom behavior.

Sending Notifications

Notification functions deliver to the Session serving the caller, so they work from any callback: the Session's own (init/2, handle_info/2, ...) and the request handlers of tools, prompts and resources, which run in a task the Session starts. From any other process, use send/2 with the session PID.

# Inside a callback:
def handle_info(:data_changed, frame) do
  Anubis.Server.send_tools_list_changed()
  {:noreply, frame}
end

Summary

Callbacks

Returns the caching hints for a cacheable stateless-era result.

Handles incoming MCP notifications from clients.

Handles a prompt get request.

Low-level handler for any MCP request.

Handles a resource read request.

Called when a session is being auto-recovered after expiry.

Handles a tool call request.

Called after a client requests a initialize request.

Transforms frame.assigns into a JSON-safe map before session persistence.

Returns optional instructions describing how to use the server and its features.

Returns the instructions for this connection, given the frame the handshake will answer with.

Returns the tools this connection may list and call, given its frame.

Returns the protocol versions served by this server.

Functions

Registers a component (tool, prompt, or resource) with the server.

Sends a notifications/elicitation/complete notification.

Sends an elicitation/create request to the client.

Sends a log message to the client, as a notifications/message.

Sends a progress notification for an ongoing operation.

Sends a prompts list changed notification.

Sends a resource updated notification for a specific resource.

Sends a resources list changed notification.

Sends a roots/list request to the client.

Sends a sampling/createMessage request to the client.

Sends a notifications/tasks/status notification with the current state of the given task.

Sends a tools list changed notification.

Sends a URL-mode elicitation/create request to the client.

Types

mcp_error()

@type mcp_error() :: Anubis.MCP.Error.t()

notification()

@type notification() :: map()

progress_step()

@type progress_step() :: number()

progress_token()

@type progress_token() :: String.t() | non_neg_integer()

progress_total()

@type progress_total() :: number()

request()

@type request() :: map()

response()

@type response() :: map()

server_capabilities()

@type server_capabilities() :: map()

server_info()

@type server_info() :: map()

Callbacks

cache_hints(method)

(optional)
@callback cache_hints(method :: String.t()) :: %{
  ttl_ms: non_neg_integer(),
  scope: :public | :private
}

Returns the caching hints for a cacheable stateless-era result.

Revision 2026-07-28 requires ttlMs and cacheScope on the complete results of server/discover, tools/list, prompts/list, resources/list, resources/templates/list and resources/read. Optional; without it every such result says ttl_ms: 0 (stale at once) and scope: :private.

:public means the result is the same for every caller, so a shared cache may serve it to anyone; answer :private for anything that varies with the caller, such as a tool list filtered per user. A handler that puts "ttlMs" or "cacheScope" in its own result keeps them. Results of a multi round-trip retry, which carry client input, always say ttl_ms: 0.

Examples

@impl Anubis.Server
def cache_hints("tools/list"), do: %{ttl_ms: to_timeout(minute: 5), scope: :private}
def cache_hints(_method), do: %{ttl_ms: 0, scope: :private}

handle_call(request, from, t)

(optional)
@callback handle_call(
  request :: term(),
  from :: GenServer.from(),
  Anubis.Server.Frame.t()
) ::
  {:reply, reply :: term(), Anubis.Server.Frame.t()}
  | {:reply, reply :: term(), Anubis.Server.Frame.t(),
     timeout() | :hibernate | {:continue, arg :: term()}}
  | {:noreply, Anubis.Server.Frame.t()}
  | {:noreply, Anubis.Server.Frame.t(),
     timeout() | :hibernate | {:continue, arg :: term()}}
  | {:stop, reason :: term(), reply :: term(), Anubis.Server.Frame.t()}
  | {:stop, reason :: term(), Anubis.Server.Frame.t()}

handle_cast(request, t)

(optional)
@callback handle_cast(request :: term(), Anubis.Server.Frame.t()) ::
  {:noreply, Anubis.Server.Frame.t()}
  | {:noreply, Anubis.Server.Frame.t(),
     timeout() | :hibernate | {:continue, arg :: term()}}
  | {:stop, reason :: term(), Anubis.Server.Frame.t()}

handle_completion(ref, argument, t)

(optional)
@callback handle_completion(ref :: String.t(), argument :: map(), Anubis.Server.Frame.t()) ::
  {:reply, Anubis.Server.Response.t() | map(), Anubis.Server.Frame.t()}
  | {:error, mcp_error(), Anubis.Server.Frame.t()}

handle_elicitation(response, request_id, t)

(optional)
@callback handle_elicitation(
  response :: map(),
  request_id :: String.t(),
  Anubis.Server.Frame.t()
) ::
  {:noreply, Anubis.Server.Frame.t()}
  | {:stop, reason :: term(), Anubis.Server.Frame.t()}

handle_info(event, t)

(optional)
@callback handle_info(event :: term(), Anubis.Server.Frame.t()) ::
  {:noreply, Anubis.Server.Frame.t()}
  | {:noreply, Anubis.Server.Frame.t(),
     timeout() | :hibernate | {:continue, arg :: term()}}
  | {:stop, reason :: term(), Anubis.Server.Frame.t()}

handle_notification(notification, state)

(optional)
@callback handle_notification(
  notification :: notification(),
  state :: Anubis.Server.Frame.t()
) ::
  {:noreply, new_state :: Anubis.Server.Frame.t()}
  | {:error, error :: mcp_error(), new_state :: Anubis.Server.Frame.t()}

Handles incoming MCP notifications from clients.

handle_prompt_get(name, arguments, t)

(optional)
@callback handle_prompt_get(
  name :: String.t(),
  arguments :: map(),
  Anubis.Server.Frame.t()
) ::
  {:reply, messages :: list(), Anubis.Server.Frame.t()}
  | {:error, mcp_error(), Anubis.Server.Frame.t()}

Handles a prompt get request.

handle_request(request, state)

(optional)
@callback handle_request(request :: request(), state :: Anubis.Server.Frame.t()) ::
  {:reply, response :: response(), new_state :: Anubis.Server.Frame.t()}
  | {:noreply, new_state :: Anubis.Server.Frame.t()}
  | {:error, error :: mcp_error(), new_state :: Anubis.Server.Frame.t()}

Low-level handler for any MCP request.

When implemented, it bypasses automatic routing to specific handlers. Return {:reply, response, frame} to answer, {:error, error, frame} to report a failure, or {:noreply, frame} to complete without a response body. A no-reply completion is not deferred work: stateless HTTP returns 202 if headers have not been sent, or closes an open SSE response without a final event.

handle_resource_read(uri, t)

(optional)
@callback handle_resource_read(uri :: String.t(), Anubis.Server.Frame.t()) ::
  {:reply, content :: map(), Anubis.Server.Frame.t()}
  | {:error, mcp_error(), Anubis.Server.Frame.t()}

Handles a resource read request.

handle_roots(roots, request_id, t)

(optional)
@callback handle_roots(
  roots :: [map()],
  request_id :: String.t(),
  Anubis.Server.Frame.t()
) ::
  {:noreply, Anubis.Server.Frame.t()}
  | {:stop, reason :: term(), Anubis.Server.Frame.t()}

handle_sampling(response, request_id, t)

(optional)
@callback handle_sampling(
  response :: map(),
  request_id :: String.t(),
  Anubis.Server.Frame.t()
) ::
  {:noreply, Anubis.Server.Frame.t()}
  | {:stop, reason :: term(), Anubis.Server.Frame.t()}

handle_session_expired(session_id, t)

(optional)
@callback handle_session_expired(session_id :: String.t(), Anubis.Server.Frame.t()) ::
  {:ok, Anubis.Server.Frame.t()}
  | {:ok, client_info :: map(), Anubis.Server.Frame.t()}
  | {:error, reason :: term()}

Called when a session is being auto-recovered after expiry.

Invoked during auto_initialize/1 instead of the normal client handshake. Receives the session ID and the current frame (pre-populated from the session store if one is configured).

Return values:

  • {:ok, frame} — accept recovery using synthetic client info
  • {:ok, client_info, frame} — accept recovery and supply real client info
  • {:error, reason} — reject recovery; the client receives an internal error

If this callback is not implemented, the default behavior is unchanged: synthetic client info is used and init/2 is called normally.

handle_tool_call(name, arguments, t)

(optional)
@callback handle_tool_call(
  name :: String.t(),
  arguments :: map(),
  Anubis.Server.Frame.t()
) ::
  {:reply, result :: term(), Anubis.Server.Frame.t()}
  | {:error, mcp_error(), Anubis.Server.Frame.t()}

Handles a tool call request.

This callback is invoked when a client calls a specific tool. It receives the tool name, the arguments provided by the client, and the current frame.

init(client_info, t)

(optional)
@callback init(client_info :: map(), Anubis.Server.Frame.t()) ::
  {:ok, Anubis.Server.Frame.t()} | {:error, term()}

Called after a client requests a initialize request.

This callback is invoked while the MCP handshake starts and so the client may not sent the notifications/initialized message yet. For checking if the notification was already sent and the MCP handshake was successfully completed, you can check the context.initialized field in the frame.

It receives the client's information and the current frame, allowing you to perform client-specific setup, validate capabilities, or prepare resources based on the connected client.

Returning {:error, reason} rejects the session: the transport answers the request with an encodable JSON-RPC error. Return an Anubis.MCP.Error struct to control the error code sent to the client; any other term is wrapped as an internal error with a stringified message.

Requests of the stateless era (2026-07-28 onward) have no handshake. Over Streamable HTTP each one is served by a session of its own, and init/2 runs for every such request, with the client info that request carries, before it is handled; clientInfo is optional in that era, and a request without it passes an empty map. The frame it returns lives only as long as the request.

serialize_assigns(assigns)

(optional)
@callback serialize_assigns(assigns :: map()) :: map()

Transforms frame.assigns into a JSON-safe map before session persistence.

frame.assigns is copied verbatim into the session store. Runtime values that are not JSON-encodable (Ecto structs, MapSet, PIDs, functions, etc.) make persistence fail. Implement this callback to project assigns down to the JSON-safe subset you actually need persisted (e.g. %{"user_id" => id}), avoiding storing full domain structs or PII.

On recovery, the restored frame is handed back to handle_session_expired/2, where the host rehydrates runtime state from these persisted values.

Defaults to the identity function when not implemented.

server_capabilities()

@callback server_capabilities() :: server_capabilities()

server_info()

@callback server_info() :: server_info()

server_instructions()

(optional)
@callback server_instructions() :: String.t() | nil

Returns optional instructions describing how to use the server and its features.

This can be used by clients to improve the LLM's understanding of available tools, resources, etc. It can be thought of like a "hint" to the model. For example, this information MAY be added to the system prompt.

Return nil to omit the instructions field from the initialize response.

server_instructions(frame)

(optional)
@callback server_instructions(frame :: Anubis.Server.Frame.t()) :: String.t() | nil

Returns the instructions for this connection, given the frame the handshake will answer with.

Optional, and takes precedence over server_instructions/0 when defined. The frame carries the transport's assigns — whatever a Plug put on the connection before the request reached the server — so a server whose guidance depends on who is asking can vary it here.

init/2 cannot serve this purpose: it runs when the client acknowledges the handshake, after the initialize result carrying the instructions has already been sent.

Examples

@impl Anubis.Server
def server_instructions(frame) do
  case frame.assigns[:plan] do
    :enterprise -> "You may call `bulk_export`. Confirm the range before exporting."
    _ -> "Exports run one record at a time."
  end
end

Return nil to omit the field from the initialize result.

server_tools(tools, frame)

(optional)
@callback server_tools(
  tools :: [Anubis.Server.Component.Tool.t()],
  frame :: Anubis.Server.Frame.t()
) :: [
  Anubis.Server.Component.Tool.t()
]

Returns the tools this connection may list and call, given its frame.

Optional. It is applied to the compile-time components concatenated with the frame's runtime tools, every time the list is built — which is both tools/list and the lookup tools/call does, so a tool withheld here cannot be reached by name either.

The frame carries the transport's assigns, so a server whose surface depends on who is asking resolves it here rather than at registration. init/2 cannot serve this purpose: it runs once per session, and on a session restored from a store it runs before the request's assigns are attached.

Return the tools unchanged to serve everyone the same surface. A tool may also be returned rewritten — a description or a schema that names something one caller may not be shown is a connection-level decision, and the struct is the place to make it.

Examples

@impl Anubis.Server
def server_tools(tools, frame) do
  if frame.assigns[:plan] == :enterprise do
    tools
  else
    Enum.reject(tools, &(&1.name == "bulk_export"))
  end
end

supported_protocol_versions()

@callback supported_protocol_versions() :: [String.t()]

Returns the protocol versions served by this server.

Defaults to all registered versions, including the stateless era. Set the :protocol_versions use option or implement this callback to restrict them. The initialize handshake only negotiates legacy versions.

terminate(reason, t)

(optional)
@callback terminate(reason :: term(), Anubis.Server.Frame.t()) :: term()

Functions

component(module, opts \\ [])

(macro)

Registers a component (tool, prompt, or resource) with the server.

notify_elicitation_complete(elicitation_id)

@spec notify_elicitation_complete(String.t()) :: :ok

Sends a notifications/elicitation/complete notification.

Tells the client that the out-of-band interaction started by a URL-mode elicitation finished. Delivery is not guaranteed, so clients must not block on it, but it lets them react programmatically.

send_elicitation_request(message, requested_schema, opts \\ [])

@spec send_elicitation_request(String.t(), map(), configuration) ::
  :ok | {:error, term()}
when configuration: [{:timeout, non_neg_integer() | nil}]

Sends an elicitation/create request to the client.

Per the MCP 2025-06-18 specification, the server provides a human-readable message and a restricted-subset JSON requested_schema describing the expected user input. The client presents this to the user and returns one of three actions: accept (with content matching the schema), decline, or cancel.

This is an asynchronous operation. The response will be delivered to your handle_elicitation/3 callback.

The requested_schema is validated synchronously before any wire I/O. The client must advertise the elicitation capability or the call returns {:error, :capability_not_supported} after enqueueing.

Schema Subset

  • Top level must be %{"type" => "object", "properties" => %{...}}
  • Properties may declare "type" of "string", "number", "integer", "boolean", or use "enum" (string-only)
  • String properties may set "format" of "email", "uri", "date", or "date-time"

Per the spec, servers MUST NOT request sensitive information through elicitation.

Example

defmodule MyServer.Tools.Greet do
  use Anubis.Server.Component, type: :tool

  @impl true
  def execute(_args, frame) do
    Anubis.Server.send_elicitation_request("What's your name?", %{
      "type" => "object",
      "properties" => %{
        "name" => %{"type" => "string", "minLength" => 1}
      },
      "required" => ["name"]
    })

    {:reply, "asked for name", frame}
  end
end

send_log_message(level, message, data \\ nil)

@spec send_log_message(
  level :: Logger.level(),
  message :: String.t(),
  metadata :: term()
) :: :ok

Sends a log message to the client, as a notifications/message.

The notification carries level and data. data is message alone, or %{"message" => message, "data" => data} when data is given. Messages below the client's log threshold are dropped. Stateless requests receive logs only when they declare io.modelcontextprotocol/logLevel; HTTP also requires SSE. Legacy sessions without a configured threshold receive all levels.

Returns :ok after enqueueing; this does not acknowledge client delivery.

Must be called from a callback — see send_resources_list_changed/0 for details.

send_progress(progress_token, progress, opts \\ [])

@spec send_progress(progress_token(), progress_step(), opts) :: :ok
when opts: [total: progress_total(), message: String.t()]

Sends a progress notification for an ongoing operation.

Pass the client's token from Anubis.Server.Frame.progress_token/1, the current progress value, and optional :total and :message fields. Call only when the request supplies a token. The notification is queued for the session serving the callback; :ok does not acknowledge client delivery. Stateless HTTP delivers progress on the request's response when it accepts SSE.

send_prompts_list_changed()

@spec send_prompts_list_changed() :: :ok

Sends a prompts list changed notification.

Must be called from a callback — see send_resources_list_changed/0 for details.

send_resource_updated(uri, timestamp \\ nil)

@spec send_resource_updated(uri :: String.t(), timestamp :: DateTime.t() | nil) :: :ok

Sends a resource updated notification for a specific resource.

Subscription-gated: only emits if the current session has previously received a resources/subscribe request for this URI. Calls for unsubscribed URIs are silently dropped.

Must be called from a callback — see send_resources_list_changed/0 for details.

send_resources_list_changed()

@spec send_resources_list_changed() :: :ok

Sends a resources list changed notification.

Must be called from a callback — a Session callback or a tool, prompt or resource handler. From any other process the message is lost.

For external processes, use send(session_pid, {:send_notification, "notifications/resources/list_changed", %{}}).

send_roots_request(opts \\ [])

@spec send_roots_request([{:timeout, non_neg_integer() | nil}]) :: :ok

Sends a roots/list request to the client.

send_sampling_request(messages, opts \\ [])

@spec send_sampling_request([map()], configuration) :: :ok
when configuration: [
       model_preferences: map() | nil,
       system_prompt: String.t() | nil,
       max_tokens: non_neg_integer() | nil,
       timeout: non_neg_integer() | nil
     ]

Sends a sampling/createMessage request to the client.

This is an asynchronous operation. The response will be delivered to your handle_sampling/3 callback.

send_task_status(task_id)

@spec send_task_status(task_id :: String.t()) :: :ok

Sends a notifications/tasks/status notification with the current state of the given task.

Must be called from a callback — see send_resources_list_changed/0 for details.

Per spec (2025-11-25), receivers MAY send these notifications when a task's status changes; they are optional and requestors MUST NOT rely on them. This helper looks up the task in the configured task store and emits the full Task projection.

send_tools_list_changed()

@spec send_tools_list_changed() :: :ok

Sends a tools list changed notification.

Must be called from a callback — see send_resources_list_changed/0 for details.

send_url_elicitation_request(url, elicitation_id, message, opts \\ [])

@spec send_url_elicitation_request(String.t(), String.t(), String.t(), configuration) ::
  :ok | {:error, term()}
when configuration: [{:timeout, non_neg_integer() | nil}]

Sends a URL-mode elicitation/create request to the client.

URL mode is the out-of-band form of elicitation: the user completes the interaction in their browser and the client never sees the sensitive data. That makes it the right mode for credentials, third-party OAuth and payment flows, all of which in-band form elicitation must never carry.

Unlike the form-mode call, there is no requested_schema: the parameters are url, the opaque elicitation_id, and message. The client answers with an action and no content.

The client must declare the url elicitation mode (%{"elicitation" => %{"url" => %{}}}). A client that declares only elicitation: %{} is form-only by specification, so the request is dropped as a missing-capability failure.

After the interaction finishes, send notify_elicitation_complete/1.

Example

Anubis.Server.send_url_elicitation_request(
  "https://github.com/login/oauth/authorize?client_id=abc",
  "550e8400-e29b-41d4-a716-446655440000",
  "Authorize access to your GitHub repositories"
)