microsoft · lesandiz · Apr 21, 2026 · Apr 21, 2026
diff --git a/.claude/skills/conductor/SKILL.md b/.claude/skills/conductor/SKILL.md
@@ -5,7 +5,7 @@ description: Validate, run, and execute workflows; creating new workflows when e
 
 # Conductor
 
-CLI tool for defining and running multi-agent workflows with the GitHub Copilot SDK or Anthropic Claude.
+CLI tool for defining and running multi-agent workflows with the GitHub Copilot SDK, Anthropic Claude, or Claude Agent SDK.
 
 > **DO NOT create new workflow files unless the user explicitly asks you to create one.** Default to running, validating, or debugging existing workflows. If the user's request is ambiguous, assume they want to run or modify an existing workflow rather than create a new one.
 

diff --git a/.claude/skills/conductor/references/yaml-schema.md b/.claude/skills/conductor/references/yaml-schema.md
@@ -6,7 +6,7 @@ Complete reference for all YAML configuration options. Derived from the Pydantic
 
 ```yaml
 workflow: WorkflowDef              # Required: workflow configuration
-tools: [string]                    # Optional: workflow-level tool names
+tools: [string]                    # Optional: workflow-level tool names (ignored by claude-agent-sdk — uses CLI config)
 agents: [AgentDef]                 # Required: agent definitions
 parallel: [ParallelGroup]         # Optional: static parallel groups
 for_each: [ForEachDef]            # Optional: dynamic parallel groups
@@ -27,14 +27,14 @@ workflow:
 
   # Runtime configuration
   runtime:
-    provider: string                # "copilot" (default) or "claude"
+    provider: string                # "copilot" (default), "claude", or "claude-agent-sdk"
     default_model: string           # Default model for all agents
-    temperature: float              # 0.0-1.0, controls randomness (optional)
-    max_tokens: integer             # Max OUTPUT tokens per response, 1-200000 (optional)
-    timeout: float                  # Per-request timeout in seconds (optional, default: 600)
+    temperature: float              # 0.0-1.0, controls randomness (optional, copilot/claude only)
+    max_tokens: integer             # Max OUTPUT tokens per response, 1-200000 (optional, copilot/claude only)
+    timeout: float                  # Per-request timeout in seconds (optional, default: 600, copilot/claude only)
     max_agent_iterations: integer   # Max tool-use roundtrips per agent (1-500, optional)
     max_session_seconds: float      # Wall-clock timeout per agent session in seconds (optional)
-    mcp_servers:                    # MCP server configurations
+    mcp_servers:                    # MCP server configurations (ignored by claude-agent-sdk — uses CLI config)
       <server_name>:
         type: string                # "stdio" (default), "http", or "sse"
         command: string             # Command to run (required for stdio)
@@ -93,7 +93,7 @@ agents:
     type: string                    # "agent" (default), "human_gate", "script", or "workflow"
     description: string             # What this agent does
     model: string                   # Override default_model
-    provider: string                # Per-agent provider override ("copilot" or "claude")
+    provider: string                # Per-agent provider override ("copilot", "claude", or "claude-agent-sdk")
 
     # Input specification (for explicit context mode)
     input:

diff --git a/AGENTS.md b/AGENTS.md
@@ -88,6 +88,7 @@ make validate-examples    # validate all examples
   - `base.py` - `AgentProvider` ABC defining `execute()`, `validate_connection()`, `close()`
   - `copilot.py` - GitHub Copilot SDK implementation
   - `claude.py` - Anthropic Claude API implementation
+  - `claude_agent_sdk.py` - Claude Agent SDK implementation (uses `claude-agent-sdk` package)
   - `factory.py` - Provider instantiation
 
 - **gates/**: Human-in-the-loop support
@@ -146,7 +147,7 @@ Use `pytest.mark.performance` for performance tests (exclude with `-m "not perfo
 
 ### Provider Parity
 
-All providers (`copilot.py`, `claude.py`) must maintain feature parity. Any change to one provider's behavior, contract, or capabilities must be applied to all providers. This includes:
+All providers must maintain feature parity where applicable. Any change to one provider's behavior, contract, or capabilities must be applied to all providers. This includes:
 
 - **Event callbacks**: Same event types emitted at the same semantic points
   - `agent_turn_start` with `{"turn": "awaiting_model"}` — immediately before each API call
@@ -160,3 +161,11 @@ All providers (`copilot.py`, `claude.py`) must maintain feature parity. Any chan
 - **Session management**: Same lifecycle (`validate_connection()`, `execute()`, `close()`)
 
 When modifying any provider, check all other providers for the same change. The dashboard, JSONL logger, console subscriber, and workflow engine all depend on consistent behavior across providers.
+
+#### `claude_agent_sdk.py` parity notes
+
+The Claude Agent SDK provider (`claude_agent_sdk.py`) delegates the agentic loop to the `claude` CLI via the `claude-agent-sdk` package. This achieves **event and output parity** but the following are managed by the SDK rather than Conductor:
+
+- **Retry and error handling**: The SDK handles retries, backoff, and parse recovery internally. The provider wraps SDK errors in `ProviderError` but does not implement its own retry logic.
+- **Tool execution**: Tools and MCP servers are managed by the `claude` CLI's own configuration. Workflow-level `tools` and `runtime.mcp_servers` fields are ignored.
+- **Runtime config**: `temperature`, `max_tokens`, and `timeout` are not configurable per-workflow — they are controlled by the CLI.
diff --git a/README.md b/README.md
@@ -14,7 +14,7 @@ Conductor provides the patterns that work: evaluator-optimizer loops for iterati
 ## Features
 
 - **YAML-based workflows** - Define multi-agent workflows in readable YAML
-- **Multiple providers** - GitHub Copilot or Anthropic Claude with seamless switching
+- **Multiple providers** - GitHub Copilot, Anthropic Claude, or Claude Agent SDK with seamless switching
 - **Parallel execution** - Run agents concurrently (static groups or dynamic for-each)
 - **Script steps** - Run shell commands and route on exit code without an AI agent
 - **Conditional routing** - Route between agents based on output conditions
@@ -155,13 +155,13 @@ conductor stop
 
 Conductor supports multiple AI providers. Choose based on your needs:
 
-| Feature | Copilot | Claude |
-|---------|---------|--------|
-| **Pricing** | Subscription ($10-39/mo) | Pay-per-token |
-| **Context Window** | 8K-128K tokens | 200K tokens |
-| **Tool Support (MCP)** | Yes | Planned |
-| **Streaming** | Yes | Planned |
-| **Best For** | Heavy usage, tools | Large context, pay-per-use |
+| Feature | Copilot | Claude | Claude Agent SDK |
+|---------|---------|--------|------------------|
+| **Pricing** | Subscription ($10-39/mo) | Pay-per-token | Via Claude Code CLI |
+| **Context Window** | 8K-128K tokens | 200K tokens | 200K tokens |
+| **Tool Support (MCP)** | Yes | Planned | Yes (built-in) |
+| **Streaming** | Yes | Planned | Yes |
+| **Best For** | Heavy usage, tools | Large context, pay-per-use | Full Claude Code toolset |
 
 ### Using Claude
 
@@ -174,6 +174,19 @@ workflow:
 
 Set your API key: `export ANTHROPIC_API_KEY=sk-ant-...`
 
+### Using Claude Agent SDK
+
+```yaml
+workflow:
+  runtime:
+    provider: claude-agent-sdk
+    default_model: claude-sonnet-4-6
+```
+
+Requires the `claude` CLI to be installed and authenticated. Install the SDK: `uv add claude-agent-sdk`
+
+> **Note:** The `claude-agent-sdk` provider delegates tool and MCP server management to the `claude` CLI. Workflow-level `tools` and `runtime.mcp_servers` fields are ignored — configure these through your Claude Code settings instead.
+
 **See also:** [Claude Documentation](docs/providers/claude.md) | [Provider Comparison](docs/providers/comparison.md) | [Migration Guide](docs/providers/migration.md)
 
 ## CLI Reference

diff --git a/docs/cli-reference.md b/docs/cli-reference.md
@@ -23,7 +23,7 @@ conductor run <workflow.yaml> [OPTIONS]
 |--------|-------|-------------|
 | `--input NAME=VALUE` | `-i` | Workflow input (repeatable) |
 | `--input.NAME=VALUE` | | Alternative input syntax |
-| `--provider PROVIDER` | `-p` | Override provider (copilot, claude) |
+| `--provider PROVIDER` | `-p` | Override provider (copilot, claude, claude-agent-sdk) |
 | `--dry-run` | | Show execution plan without running |
 | `--skip-gates` | | Auto-select first option at human gates |
 | `--quiet` | `-q` | Minimal output (agent lifecycle and routing only) |

diff --git a/docs/configuration.md b/docs/configuration.md
@@ -11,7 +11,7 @@ The `runtime` section of your workflow defines provider settings and global defa
 ```yaml
 workflow:
   runtime:
-    provider: copilot  # or 'claude'
+    provider: copilot  # or 'claude' or 'claude-agent-sdk'
     default_model: gpt-5.2
     # Provider-specific settings...
 ```

diff --git a/docs/providers/comparison.md b/docs/providers/comparison.md
@@ -1,21 +1,21 @@
-# Provider Comparison: Copilot vs Claude
+# Provider Comparison: Copilot vs Claude vs Claude Agent SDK
 
-This guide helps you choose between GitHub Copilot and Anthropic Claude providers for your workflows.
+This guide helps you choose between GitHub Copilot, Anthropic Claude, and Claude Agent SDK providers for your workflows.
 
 ## Quick Comparison
 
-| Feature | Copilot | Claude | Winner |
-|---------|---------|--------|--------|
-| **Context Window** | 8K-128K | 200K (all models) | Claude |
-| **Pricing Model** | Subscription ($10-39/mo) | Pay-per-token | Depends |
-| **Setup** | GitHub auth | API key | Copilot (easier) |
-| **Model Selection** | GPT-5.2, o1 | Haiku, Sonnet, Opus | Tie |
-| **Streaming** | Yes | No (Phase 1) | Copilot |
-| **Tool Support** | Yes (MCP) | No (Phase 1) | Copilot |
-| **Speed** | Fast | Fast | Tie |
-| **Output Quality** | Excellent | Excellent | Tie |
-| **Cost Predictability** | High (flat rate) | Variable (usage-based) | Copilot |
-| **Multi-provider** | No | Yes (via Conductor) | Claude |
+| Feature | Copilot | Claude | Claude Agent SDK |
+|---------|---------|--------|------------------|
+| **Context Window** | 8K-128K | 200K (all models) | 200K |
+| **Pricing Model** | Subscription ($10-39/mo) | Pay-per-token | Via Claude Code CLI |
+| **Setup** | GitHub auth | API key | `claude` CLI auth |
+| **Model Selection** | GPT-5.2, o1 | Haiku, Sonnet, Opus | Haiku, Sonnet, Opus |
+| **Streaming** | Yes | No (Phase 1) | Yes |
+| **Tool Support** | Yes (MCP) | No (Phase 1) | Yes (built-in) |
+| **Speed** | Fast | Fast | Fast |
+| **Output Quality** | Excellent | Excellent | Excellent |
+| **Cost Predictability** | High (flat rate) | Variable (usage-based) | Variable |
+| **Agentic Loop** | SDK-managed | Manual (provider code) | SDK-managed |
 
 ## When to Use Copilot
 
@@ -111,6 +111,51 @@ agents:
     prompt: "Analyze the following document ({{ document | length }} chars)"
 ```
 
+## When to Use Claude Agent SDK
+
+### ✅ Choose Claude Agent SDK if:
+
+1. **You want built-in tool support with Claude models**
+   - WebSearch, WebFetch, Bash, file operations out of the box
+   - No MCP server configuration needed for common tools
+   - Full Claude Code toolset available
+
+2. **You already use the `claude` CLI**
+   - Authentication handled by the CLI
+   - No separate API key management
+   - Settings inherited from your Claude Code environment
+
+3. **You want the SDK to manage the agentic loop**
+   - Retry logic, tool execution, and structured output handled by the SDK
+   - Less provider code to maintain
+   - Native interrupt support
+
+4. **You need streaming with Claude models**
+   - Real-time message streaming (unlike the raw Claude provider)
+   - Typed message objects for each event
+
+### Important: Tools and MCP Servers
+
+The `claude-agent-sdk` provider delegates tool and MCP server management entirely to the `claude` CLI. This means:
+
+- Workflow-level `tools` and `runtime.mcp_servers` fields are **ignored** — configure tools and MCP servers through your Claude Code settings instead
+- The full Claude Code toolset (WebSearch, Bash, Read, Write, etc.) is available automatically
+- `temperature`, `max_tokens`, and `timeout` are also managed by the CLI and not configurable per-workflow
+
+### Example Claude Agent SDK Workflow
+
+```yaml
+workflow:
+  name: sdk-workflow
+  runtime:
+    provider: claude-agent-sdk
+    default_model: claude-sonnet-4-6
+
+agents:
+  - name: researcher
+    prompt: "Research {{ topic }} using web search"
+```
+
 ## Cost Comparison
 
 ### Scenario 1: Light Usage (10 hours/month)
@@ -298,6 +343,8 @@ Use this matrix to decide:
 | Process long documents | **Claude** |
 | Complex reasoning tasks | **Claude** (Opus) |
 | Simple high-volume tasks | **Claude** (Haiku 4.5) |
+| Already use `claude` CLI | **Claude Agent SDK** |
+| Want streaming with Claude | **Claude Agent SDK** |
 
 ## Multi-Provider Strategy
 
@@ -347,4 +394,11 @@ workflow:
 - ✅ Long document processing
 - ✅ Cost optimization (Haiku)
 
-**Bottom line**: Both are excellent. Choose based on your usage patterns, budget, and feature requirements. Conductor makes it easy to switch between them or use both strategically.
+**Choose Claude Agent SDK** for:
+- ✅ Built-in tools (WebSearch, Bash, etc.)
+- ✅ Streaming with Claude models
+- ✅ SDK-managed agentic loop
+- ✅ Existing `claude` CLI users
+- ✅ No API key management
+
+**Bottom line**: All three are excellent. Choose based on your usage patterns, budget, and feature requirements. Conductor makes it easy to switch between them or use all three strategically.
diff --git a/examples/test-claude-agent-sdk.yaml b/examples/test-claude-agent-sdk.yaml
@@ -0,0 +1,53 @@
+# Simple Question-Answering Workflow using Claude Agent SDK
+#
+# This example demonstrates a basic linear workflow with a single agent
+# that answers questions. It shows:
+# - Basic workflow structure
+# - Input parameters
+# - Output schema validation
+# - Simple routing to $end
+#
+# Usage:
+#   conductor run examples/test-claude-agent-sdk.yaml --input question="What is Microsoft Conductor?"
+#
+# Note: Requires Claude Code CLI installed and configured. Adjust the runtime model as needed based on your Claude Code setup.
+
+workflow:
+  name: simple-qa
+  description: A simple question-answering workflow with a single agent
+  version: "1.0.0"
+  entry_point: answerer
+
+  runtime:
+    provider: claude-agent-sdk
+    default_model: claude-haiku-4-5@20251001 # Vertex AI naming; it depends on the backedend your Claude Code setup uses. Adjust as needed.
+
+  input:
+    question:
+      type: string
+      required: true
+      description: The question to answer
+
+agents:
+  - name: answerer
+    description: Answers the user's question clearly and concisely
+    prompt: |
+      You are a helpful assistant and researcher. Please answer the following question 
+      clearly and concisely:
+
+      Question: {{ workflow.input.question }}
+
+      Provide a direct answer without unnecessary preamble.
+    output:
+      answer:
+        type: string
+        description: The answer to the question
+      confidence:
+        type: string
+        description: Confidence level (high, medium, low)
+    routes:
+      - to: $end
+
+output:
+  answer: "{{ answerer.output.answer }}"
+  confidence: "{{ answerer.output.confidence }}"
diff --git a/pyproject.toml b/pyproject.toml
@@ -44,6 +44,12 @@ dependencies = [
     "uvicorn>=0.30.0",
     "websockets>=12.0",
     "httpx>=0.27.0",
+    "claude-agent-sdk>=0.1.64",
+]
+
+[project.optional-dependencies]
+claude-agent-sdk = [
+    "claude-agent-sdk>=0.1.0",
 ]
 
 [project.urls]

diff --git a/src/conductor/config/schema.py b/src/conductor/config/schema.py
@@ -399,7 +399,7 @@ class AgentDef(BaseModel):
     type: Literal["agent", "human_gate", "script", "workflow"] | None = None
     """Agent type. Defaults to 'agent' if not specified."""
 
-    provider: Literal["copilot", "claude"] | None = None
+    provider: Literal["copilot", "claude", "claude-agent-sdk"] | None = None
     """Provider override for this agent.
 
     If None (default), the agent uses the workflow.runtime.provider.
@@ -630,7 +630,7 @@ def validate_type_requirements(self) -> MCPServerDef:
 class RuntimeConfig(BaseModel):
     """Provider and runtime configuration."""
 
-    provider: Literal["copilot", "openai-agents", "claude"] = "copilot"
+    provider: Literal["copilot", "openai-agents", "claude", "claude-agent-sdk"] = "copilot"
     """SDK provider to use for agent execution."""
 
     default_model: str | None = None

diff --git a/src/conductor/providers/__init__.py b/src/conductor/providers/__init__.py
@@ -6,12 +6,14 @@
 
 from conductor.providers.base import AgentOutput, AgentProvider
 from conductor.providers.claude import ClaudeProvider
+from conductor.providers.claude_agent_sdk import ClaudeAgentSdkProvider
 from conductor.providers.copilot import CopilotProvider
 from conductor.providers.factory import create_provider
 
 __all__ = [
     "AgentOutput",
     "AgentProvider",
+    "ClaudeAgentSdkProvider",
     "ClaudeProvider",
     "CopilotProvider",
     "create_provider",