Skip to main content
Claude Code supports two modes of operation:
ModeDescription
Interactive REPLDefault mode. Full terminal UI with message history, status bar, and keyboard shortcuts.
SDK / headlessNon-interactive mode. No UI. Accepts a prompt, runs the agent loop, and writes structured output to stdout.
Headless mode is designed for scripting, CI/CD pipelines, editor integrations, and any context where a human is not directly driving the session.

​
SDK entrypoint

The SDK surface lives in entrypoints/sdk/ and consists of three files:
FilePurpose
coreTypes.tsCore TypeScript types for SDK input and output
coreSchemas.tsZod schemas for validating SDK messages
controlSchemas.tsSchemas for control messages (abort, status updates)
The QueryEngine class (QueryEngine.ts) is the headless query engine that drives the agent loop without any UI rendering. It accepts messages, runs tool calls, and emits structured output events.

​
CLI flags for non-interactive use

​
-p / --print

Runs Claude Code in print mode. Claude processes the prompt, executes any needed tools, and exits. No interactive input is read from stdin. 
claude -p "Summarize the changes in the last five commits"

​
--output-format

Controls the format of Claude’s output.
Emits a single JSON object after the agent loop completes. Contains the final assistant message, token usage, cost, and session metadata.
claude -p "Run all tests and fix any failures" --output-format json
Example output:
{
  "type": "result",
  "subtype": "success",
  "result": "All 42 tests pass.",
  "session_id": "sess_01abc...",
  "cost_usd": 0.0031,
  "usage": {
    "input_tokens": 1204,
    "output_tokens": 318,
    "cache_read_input_tokens": 890,
    "cache_creation_input_tokens": 0
  }
}

​
--input-format stream-json

Accepts streaming JSON input on stdin instead of a single prompt string. Each line must be a valid JSON message. This lets you pipe structured input into Claude Code from another process. 
echo '{"type":"user","message":"List all TODO comments"}' | \
  claude --input-format stream-json --output-format stream-json

​
Running Claude Code in a CI/CD pipeline

The -p flag combined with --output-format json makes Claude Code straightforward to embed in automated pipelines: 
claude -p "Run all tests and fix any failures" --output-format json
Parse the JSON result in your pipeline script to check for success, extract the response text, or record token costs. 
result=$(claude -p "Check for security issues in src/" --output-format json)
status=$(echo "$result" | jq -r '.subtype')

if [ "$status" != "success" ]; then
  echo "Claude Code reported an error"
  exit 1
fi

​
MCP server mode

Claude Code can act as a Model Context Protocol (MCP) server, exposing its tools and agent capabilities to any MCP-compatible client: 
claude --mcp-server
The MCP entrypoint is defined in entrypoints/mcp.ts. In this mode Claude Code listens for MCP requests on stdin/stdout and responds with MCP-formatted messages. No interactive UI is started.

​
Disabling animations

In headless environments the spinner and other animations are suppressed automatically. You can also force this explicitly: 
CLAUDE_CODE_DISABLE_NONINTERACTIVE_ANIMATIONS=1 claude -p "..."
This is useful when Claude Code’s output is captured by a logging system that does not handle ANSI escape sequences well.

Build docs developers (and LLMs) love

Get started for freeTalk to us