Claude Agent SDK Logo

Claude Agent SDK for Elixir

Hex.pmHex DocsHex.pm DownloadsLicenseCILast Commit

An Elixir SDK aiming for high parity with the official claude-agent-sdk-python while treating the Claude Code CLI as the authoritative runtime contract. Build AI-powered applications with Claude using a production-ready interface for the Claude Code CLI, featuring streaming responses, lifecycle hooks, permission controls, and in-process tool execution via MCP.

Note: This SDK does not bundle the Claude Code CLI. You must install it separately (see Prerequisites).

Documentation Menu

What You Can Build

Runtime Architecture

ASM Boundary

If you enter Claude through agent_session_manager, the normalized ASM kernel still stops at provider selection, lane selection, event projection, and session/run orchestration.

The optional ASM seam for Claude lives under ASM.Extensions.ProviderSDK.Claude. That extension may:

If claude_agent_sdk is present in the dependency graph, ASM activates that namespace automatically in ASM.Extensions.ProviderSDK.available_extensions/0 and ASM.Extensions.ProviderSDK.capability_report/0. Client apps do not need to register it manually.

ASM-owned fields such as cwd, permission_mode, model, max_turns, and the transport timeout still stay in ASM config. The extension's native override bag is for Claude-native fields only.

It does not move the control family into ASM. Once you cross that seam, the real control APIs remain here:

Packaging Boundary

Phase 4 finalizes the Claude release boundary:

Schema Boundary

Zoi is now the canonical boundary-schema layer for new dynamic boundary work inside claude_agent_sdk.

Centralized Model Selection

claude_agent_sdk no longer owns active model defaulting or fallback policy. That authority lives in cli_subprocess_core.

Authoritative core surface:

Claude-side behavior:

Do not treat repo-local config snapshots as authoritative model policy. The shared core registry is the source of truth.

Claude Ollama Backend

claude_agent_sdk now supports an explicit Claude :ollama backend through the same core-owned payload path.

Example:

options =
  ClaudeAgentSDK.Options.new(
    provider_backend: :ollama,
    anthropic_base_url: "http://localhost:11434",
    external_model_overrides: %{"haiku" => "llama3.2"},
    model: "haiku"
  )

The SDK still runs the normal claude binary. The core payload resolves the actual transport model and injects the Anthropic-compatible Ollama env.

Installation

Add to your mix.exs:

def deps do
  [
    {:claude_agent_sdk, "~> 0.17.0"}
  ]
end

Then fetch dependencies:

mix deps.get

Prerequisites

Install the Claude Code CLI (requires Node.js):

npm install -g @anthropic-ai/claude-code

Verify installation:

claude --version

CLI Compatibility

Quick Start

1. Authenticate

Choose one method:

# Option A: Environment variable (recommended for CI/CD)
export ANTHROPIC_API_KEY="sk-ant-api03-..."

# Option B: OAuth token
export CLAUDE_AGENT_OAUTH_TOKEN="sk-ant-oat01-..."

# Option C: Interactive login
claude login

2. Run Your First Query

alias ClaudeAgentSDK.{ContentExtractor, Options}

# Simple query with streaming collection
ClaudeAgentSDK.query("Write a function that calculates factorial in Elixir")
|> Enum.each(fn msg ->
  case msg.type do
    :assistant -> IO.puts(ContentExtractor.extract_text(msg) || "")
    :result -> IO.puts("Done! Cost: $#{msg.data.total_cost_usd}")
    _ -> :ok
  end
end)

3. Real-Time Streaming

alias ClaudeAgentSDK.Streaming

{:ok, session} = Streaming.start_session()

Streaming.send_message(session, "Explain GenServers in one paragraph")
|> Stream.each(fn
  %{type: :text_delta, text: chunk} -> IO.write(chunk)
  %{type: :message_stop} -> IO.puts("")
  _ -> :ok
end)
|> Stream.run()

Streaming.close_session(session)

If session initialization or message send fails, the stream now emits an immediate %{type: :error, error: reason} event instead of waiting for the 5-minute stream timeout.

Authentication

The SDK supports three authentication methods, checked in this order:

Method Environment Variable Best For
OAuth Token CLAUDE_AGENT_OAUTH_TOKEN Production / CI
API Key ANTHROPIC_API_KEY Development
CLI Login (uses claude login session) Local development

Cloud Providers

AWS Bedrock:

export CLAUDE_AGENT_USE_BEDROCK=1
export AWS_ACCESS_KEY_ID=...
export AWS_SECRET_ACCESS_KEY=...
export AWS_REGION=us-west-2

Google Vertex AI:

export CLAUDE_AGENT_USE_VERTEX=1
export GOOGLE_APPLICATION_CREDENTIALS=/path/to/key.json
export GOOGLE_CLOUD_PROJECT=your-project-id

Token Setup (Local Development)

For persistent authentication without re-login:

mix claude.setup_token

AuthManager keeps running if token storage save/clear fails and returns {:error, reason}. Handle clear_auth/0 accordingly in your app code:

case ClaudeAgentSDK.AuthManager.clear_auth() do
  :ok -> :ok
  {:error, reason} -> IO.puts("Failed to clear auth: #{inspect(reason)}")
end

Check authentication status:

alias ClaudeAgentSDK.AuthChecker
diagnosis = AuthChecker.diagnose()
# => %{authenticated: true, auth_method: "Anthropic API", ...}

Core Concepts

Choosing the Right API

API Use Case When to Use
query/2 Simple queries Batch processing, scripts
Streaming Typewriter UX Chat interfaces, real-time output
Client Full control Multi-turn agents, tools, hooks

Query API

The simplest way to interact with Claude:

# Basic query
messages = ClaudeAgentSDK.query("What is recursion?") |> Enum.to_list()

# With options
opts = %ClaudeAgentSDK.Options{
  model: "sonnet",
  max_turns: 5,
  output_format: :stream_json
}
messages = ClaudeAgentSDK.query("Explain OTP", opts) |> Enum.to_list()

# Streamed input prompts (unidirectional)
prompts = [
  %{"type" => "user", "message" => %{"role" => "user", "content" => "Hello"}},
  %{"type" => "user", "message" => %{"role" => "user", "content" => "How are you?"}}
]

ClaudeAgentSDK.query(prompts, opts) |> Enum.to_list()

# Execution-surface routing
opts = %ClaudeAgentSDK.Options{
  execution_surface: [
    surface_kind: :ssh_exec,
    transport_options: [
      destination: "claude.example",
      ssh_user: "sdk",
      port: 22
    ]
  ]
}

ClaudeAgentSDK.query("Hello", opts) |> Enum.to_list()

# Continue a conversation
ClaudeAgentSDK.continue("Can you give an example?") |> Enum.to_list()

# Resume a specific session
ClaudeAgentSDK.resume("session-id", "What about supervision trees?") |> Enum.to_list()

Streaming API

For real-time, character-by-character output:

alias ClaudeAgentSDK.{Options, Streaming}

{:ok, session} = Streaming.start_session(%Options{model: "haiku"})

# Send messages and stream responses
Streaming.send_message(session, "Write a haiku about Elixir")
|> Enum.each(fn
  %{type: :text_delta, text: t} -> IO.write(t)
  %{type: :tool_use_start, name: n} -> IO.puts("\nUsing tool: #{n}")
  %{type: :message_stop} -> IO.puts("\n---")
  _ -> :ok
end)

# Multi-turn conversation (context preserved)
Streaming.send_message(session, "Now write one about Phoenix")
|> Enum.to_list()

Streaming.close_session(session)

Subagent Streaming: When Claude spawns subagents via the Agent tool, events include a parent_tool_use_id field to identify the source. Main agent events have nil, subagent events have the Agent tool call ID. Streaming events also include uuid, session_id, and raw_event metadata for parity with the Python SDK. Stream event wrappers require uuid and session_id (missing keys raise). See the Streaming Guide for details.

Hooks System

Intercept and control agent behavior at key lifecycle points:

alias ClaudeAgentSDK.{Client, Options}
alias ClaudeAgentSDK.Hooks.{Matcher, Output}

# Block dangerous commands
check_bash = fn input, _id, _ctx ->
  case input do
    %{"tool_name" => "Bash", "tool_input" => %{"command" => cmd}} ->
      if String.contains?(cmd, "rm -rf") do
        Output.deny("Dangerous command blocked")
      else
        Output.allow()
      end
    _ -> %{}
  end
end

opts = %Options{
  hooks: %{
    pre_tool_use: [Matcher.new("Bash", [check_bash])]
  }
}

{:ok, client} = Client.start_link(opts)

Available Hook Events (all 12 Python SDK events supported):

See the Hooks Guide for comprehensive documentation.

Supervision

Hook and permission callbacks run in async tasks. For production, add the SDK task supervisor so callback processes are supervised:

children = [
  ClaudeAgentSDK.TaskSupervisor,
  {ClaudeAgentSDK.Client, options}
]

If you use a custom supervisor name, configure the SDK to match:

children = [
  {ClaudeAgentSDK.TaskSupervisor, name: MyApp.ClaudeTaskSupervisor}
]

config :claude_agent_sdk, task_supervisor: MyApp.ClaudeTaskSupervisor

If an explicitly configured supervisor is missing at runtime, the SDK logs a warning and falls back to Task.start/1. With default settings, missing ClaudeAgentSDK.TaskSupervisor falls back silently for backward compatibility. For stricter behavior in dev/test:

config :claude_agent_sdk, task_supervisor_strict: true

In strict mode, ClaudeAgentSDK.TaskSupervisor.start_child/2 returns {:error, {:task_supervisor_unavailable, supervisor}} instead of spawning an unsupervised fallback task.

Permission System

Fine-grained control over tool execution:

alias ClaudeAgentSDK.{Options, Permission.Result}

permission_callback = fn ctx ->
  case ctx.tool_name do
    "Write" ->
      # Redirect system file writes to safe location
      if String.starts_with?(ctx.tool_input["file_path"], "/etc/") do
        safe_path = "/tmp/sandbox/" <> Path.basename(ctx.tool_input["file_path"])
        Result.allow(updated_input: %{ctx.tool_input | "file_path" => safe_path})
      else
        Result.allow()
      end
    _ ->
      Result.allow()
  end
end

opts = %Options{
  can_use_tool: permission_callback,
  permission_mode: :default  # :default | :accept_edits | :plan | :bypass_permissions | :auto | :dont_ask
}

Note: can_use_tool is mutually exclusive with permission_prompt_tool. The SDK routes can_use_tool through the control client (including string prompts), auto-enables include_partial_messages, and sets permission_prompt_tool to \"stdio\" internally so the CLI can emit permission callbacks. Use :default, :plan, or the CLI's :auto mode for built-in tool permissions. Hook-based fallback only applies when the CLI does not emit can_use_tool, and that fallback ignores updated_permissions. :delegate is no longer forwarded because current Claude CLI builds reject it.

CLI transcript history now matches the official SDK surface:

# Claude CLI transcript history
sessions = ClaudeAgentSDK.list_sessions(directory: "/path/to/project")
messages = ClaudeAgentSDK.get_session_messages("550e8400-e29b-41d4-a716-446655440000",
  directory: "/path/to/project"
)

# SDK-managed SessionStore history
{:ok, saved_sessions} = ClaudeAgentSDK.list_saved_sessions(storage_dir: "/custom/path")

Stream a single client response until the final result:

Client.receive_response_stream(client)
|> Enum.to_list()

MCP Tools (In-Process)

Define custom tools that Claude can call directly in your application:

defmodule MyTools do
  use ClaudeAgentSDK.Tool

  deftool :calculate, "Perform a calculation", %{
    type: "object",
    properties: %{
      expression: %{type: "string", description: "Math expression to evaluate"}
    },
    required: ["expression"]
  } do
    def execute(%{"expression" => expr}) do
      # Your logic here
      result = eval_expression(expr)
      {:ok, %{"content" => [%{"type" => "text", "text" => "Result: #{result}"}]}}
    end
  end
end

# Create an MCP server with your tools
server = ClaudeAgentSDK.create_sdk_mcp_server(
  name: "calculator",
  version: "1.0.0",
  tools: [MyTools.Calculate]
)

# Without :supervisor, the SDK keeps the registry under its internal
# SDK MCP supervisor so the server survives creator process exits.

# Optional: start tool registry under your DynamicSupervisor
{:ok, sup} = DynamicSupervisor.start_link(strategy: :one_for_one)

server = ClaudeAgentSDK.create_sdk_mcp_server(
  name: "calculator",
  version: "1.0.0",
  tools: [MyTools.Calculate],
  supervisor: sup
)

opts = %ClaudeAgentSDK.Options{
  mcp_servers: %{"calc" => server},
  allowed_tools: ["mcp__calc__calculate"]
}

Note: MCP server routing only supports initialize, tools/list, tools/call, and notifications/initialized. Calls to resources/list or prompts/list return JSON-RPC method-not-found errors to match the Python SDK. If version is omitted, it defaults to "1.0.0".

Configuration Options

Key options for ClaudeAgentSDK.Options:

Option Type Description
model string "sonnet", "opus", "haiku"
effort atom :low, :medium, :high, :max — controls reasoning effort; :max is Opus-only; invalid values raise ArgumentError (not supported for Haiku)
thinking map %{type: :adaptive}, %{type: :enabled, budget_tokens: N}, %{type: :disabled}
max_turns integer Maximum conversation turns
system_prompt string Custom system instructions
output_format atom/map :text, :json, :stream_json, or JSON schema (SDK enforces stream-json for transport; JSON schema still passed)
allowed_tools list Tools Claude can use
permission_mode atom :default, :accept_edits, :plan, :bypass_permissions, :auto, :dont_ask
hooks map Lifecycle hook callbacks
mcp_servers map or string MCP server configurations (or JSON/path alias for mcp_config)
cwd string Working directory for file operations
timeout_ms integer Command timeout (default: 75 minutes)
transport_error_mode atom :result (default) or :raise for strict transport/decode failures
max_buffer_size integer Maximum JSON buffer size (default: 1MB, overflow yields CLIJSONDecodeError)

CLI path override: set path_to_claude_code_executable or executable in Options (Python cli_path equivalent).

On SSH-backed execution_surface values, Claude is resolved on the remote host. If the binary is installed outside the remote non-login PATH, pass an explicit executable or path_to_claude_code_executable for that target.

Runtime Application Config

All tunable constants (timeouts, buffer sizes, auth paths, CLI flags, env var names, concurrency limits) are centralized in Config.* sub-modules and can be overridden per-environment:

# config/config.exs
config :claude_agent_sdk, ClaudeAgentSDK.Config.Timeouts,
  query_total_ms: 5_400_000,           # total query timeout (default: 75 min)
  tool_execution_ms: 60_000            # per-tool timeout (default: 30 s)

config :claude_agent_sdk, ClaudeAgentSDK.Config.Buffers,
  max_stdout_buffer_bytes: 2_097_152   # stdout buffer (default: 1 MB)

config :claude_agent_sdk, ClaudeAgentSDK.Config.Orchestration,
  max_concurrent: 10,                  # parallel query limit (default: 5)
  max_retries: 5                       # retry attempts (default: 3)

# Legacy flat keys still work:
config :claude_agent_sdk,
  cli_stream_module: ClaudeAgentSDK.Query.CLIStream,
  task_supervisor_strict: false

See the Configuration Internals guide for the complete reference of every tunable, its default, and override examples.

config :claude_agent_sdk, :process_module is still read as a fallback for query streaming, but it is deprecated and logs a warning once per legacy module.

SessionStore now hydrates on-disk cache in a handle_continue/2 step. Startup is faster, but list/search can be briefly incomplete immediately after boot while warmup finishes.

ExternalRuntimeTransport.Transport and Streaming.Session support startup_mode: :lazy to defer subprocess startup to handle_continue/2. Deterministic startup validation still happens before start_link returns, so missing cwd/command style failures surface immediately. Once preflight passes, lazy mode can still surface subprocess launch failures as a process exit after init.

Query-side transport errors normalize equivalent reasons to stable atoms where possible: {:command_not_found, "claude"} is treated as :cli_not_found.

SDK Logging

The SDK uses its own log level filter (default: :warning) to keep output quiet in dev. Configure via application env:

config :claude_agent_sdk, log_level: :warning  # :debug | :info | :warning | :error | :off

Option Presets

alias ClaudeAgentSDK.OptionBuilder

# Environment-based presets
OptionBuilder.build_development_options()  # Permissive, verbose
OptionBuilder.build_production_options()   # Restrictive, safe
OptionBuilder.for_environment()            # Auto-detect from Mix.env()

# Use-case presets
OptionBuilder.build_analysis_options()     # Read-only code analysis
OptionBuilder.build_chat_options()         # Simple chat, no tools
OptionBuilder.quick()                      # Fast one-off queries

# Effort and thinking helpers
OptionBuilder.with_opus()
|> OptionBuilder.with_effort(:max)
|> OptionBuilder.with_thinking(%{type: :adaptive})

Examples

The examples/ directory contains runnable demonstrations.

Mix Task Example (Start Here)

If you want to integrate Claude into your own Mix project, see the mix_task_chat example — a complete working app with Mix tasks:

cd examples/mix_task_chat
mix deps.get
mix chat "Hello, Claude!"           # Streaming response
mix chat --interactive              # Multi-turn conversation
mix ask -q "What is 2+2?"           # Script-friendly output

Script Examples

# Run all examples
bash examples/run_all.sh

# Run a specific example
mix run examples/basic_example.exs
mix run examples/streaming_tools/quick_demo.exs
mix run examples/hooks/basic_bash_blocking.exs

Key Examples:

Full Application Examples

Complete Mix applications demonstrating production-ready SDK integration patterns:

Example Description Key Features
phoenix_chat/ Real-time chat with Phoenix LiveView LiveView, Channels, streaming responses, session management
document_generation/ AI-powered Excel document generation elixlsx, natural language parsing, Mix tasks
research_agent/ Multi-agent research coordination Agent tool, subagent tracking via hooks, parallel execution
skill_invocation/ Skill tool usage and tracking Skill definitions, hook-based tracking, GenServer state
email_agent/ AI-powered email assistant SQLite storage, rule-based processing, natural language queries 
# Run Phoenix Chat
cd examples/phoenix_chat && mix deps.get && mix phx.server
# Visit http://localhost:4000

# Run Document Generation
cd examples/document_generation && mix deps.get && mix generate.demo

# Run Research Agent
cd examples/research_agent && mix deps.get && mix research "quantum computing"

# Run Skill Invocation demo
cd examples/skill_invocation && mix deps.get && mix run -e "SkillInvocation.demo()"

# Run Email Agent
cd examples/email_agent && mix deps.get && mix email.assistant "find emails from last week"

Guides

Guide Description
Getting Started Installation, authentication, and first query
Streaming Real-time streaming and typewriter effects
Hooks Lifecycle hooks for tool control
MCP Tools In-process tool definitions with MCP
Permissions Fine-grained permission controls
Configuration Complete options reference
Agents Custom agent personas
Sessions Session management and persistence
Testing Mock system and testing patterns
Error Handling Error types and recovery

Upgrading

For breaking changes and migration notes, see CHANGELOG.md.

0.12.0 breaking changes:

0.11.0 breaking changes:

0.10.0 fix (resume turn persistence):

0.9.0 breaking change (streaming):

That strict requirement applies to the common CLI streaming parser. The SDK-local control lane still accepts incomplete stream_event wrappers and surfaces missing metadata as nil, which keeps control-client transports forward-compatible without weakening the core streaming contract.

Additional Resources:

License

MIT License - see LICENSE for details.

Built with Elixir and Claude

Model Selection Contract

/home/home/p/g/n/claude_agent_sdk no longer owns active model-selection policy. Model catalog visibility, defaulting, validation, and failure semantics are owned by /home/home/p/g/n/cli_subprocess_core through CliSubprocessCore.ModelRegistry.resolve/3, CliSubprocessCore.ModelRegistry.validate/2, and CliSubprocessCore.ModelRegistry.default_model/2.

The Claude SDK now treats model values as resolved payload data and limits its responsibility to option shaping and command rendering.

Session History And Intervention

The Claude runtime lane now publishes an honest session-control contract for orchestration layers.

This is meant for recovery and operator-control flows, not blind retries. The caller should resume the same session when possible and only replay work when a real continuation handle is unavailable.