Programmatic Claude Code
Run Claude non-interactively for automation workflows.
Three Interfaces
| Interface |
Install |
Best For |
CLI (-p) |
Built-in |
Scripts, CI/CD, quick tasks |
| Python SDK |
pip install claude-agent-sdk |
Custom agents, hooks, multi-turn |
| TypeScript SDK |
npm install @anthropic-ai/claude-agent-sdk |
Type-safe apps, Node.js |
Core Patterns
One-Shot Task
claude -p "Fix TypeScript errors in src/" \
--allowedTools "Read,Edit,Bash" \
--output-format json
Session Chaining
# Start task, capture session ID
result=$(claude -p "Create auth system" --output-format json)
session_id=$(echo "$result" | jq -r '.session_id')
# Continue in same session
claude -p "Add tests for auth" --resume "$session_id"
Structured Output
claude -p "List all API endpoints" \
--output-format json \
--json-schema '{"type":"object","properties":{"endpoints":{"type":"array"}}}'
Python Multi-Turn
from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions
options = ClaudeAgentOptions(
allowed_tools=["Read", "Write", "Edit", "Bash"],
permission_mode="acceptEdits"
)
async with ClaudeSDKClient(options=options) as client:
await client.query("Build the component")
async for msg in client.receive_response():
print(msg)
await client.query("Now add tests")
async for msg in client.receive_response():
print(msg)
Key Flags
| Flag |
Purpose |
-p "prompt" |
Non-interactive mode |
--allowedTools "X,Y" |
Auto-approve tools (no prompts) |
--output-format json |
Structured output with session_id |
--json-schema '{...}' |
Enforce response format |
--resume <id> |
Continue specific session |
--continue |
Continue most recent session |
Files
README.md - Overview and quick examplesharvest-and-build.md - Website cloning automation pipeline
Use Cases
- CI/CD - Run in GitHub Actions for automated fixes
- Orchestration - Chain multiple Claude sessions
- Pipelines - Harvest → Analyse → Build → Validate workflows
- Scheduled Tasks - Daily security scans, dependency updates
diegosouzapw