Skip to main content

hackagent.router.providers.claude

Claude Code provider built on top of LiteLLM.

Claude Code is Anthropic's agentic coding CLI. It serves no HTTP endpoint of its own — the supported ways to drive it locally are the headless CLI (claude -p) or the Claude Agent SDK. LiteLLM has no built-in provider for it, so — exactly like the Google ADK provider — we register a per-instance :class:litellm.CustomLLM handler under a unique provider name. Its completion shells out to claude -p instead of making an HTTP call, so the request still flows through litellm.completion and is captured by the HackAgent tracking logger like every other provider.

This makes a locally-installed Claude Code a first-class attack target: no external bridge, no HTTP server. The only prerequisite is the claude binary being on PATH (checked at adapter construction).

ClaudeCodeConfigurationError Objects​

class ClaudeCodeConfigurationError(AdapterConfigurationError)

Claude Code adapter configuration issues (e.g. binary not found).

ClaudeCodeInteractionError Objects​

class ClaudeCodeInteractionError(AdapterInteractionError)

Errors invoking the claude CLI.

ClaudeCodeResponseParsingError Objects​

class ClaudeCodeResponseParsingError(AdapterResponseParsingError)

Errors parsing the claude -p --output-format json output.

ClaudeCodeAgent Objects​

class ClaudeCodeAgent(Agent)

Adapter for a locally-installed Claude Code CLI.

Drives Claude Code in headless mode (claude -p) through a per-instance :class:litellm.CustomLLM handler registered under a unique provider name (hackagent_claude_code_<id>), so requests flow through litellm.completion like every other provider — even though Claude Code speaks no HTTP.

Required config:

  • name: the Claude model to drive (sonnet/opus/haiku aliases or a full id like claude-opus-4-8). Used as both the --model value and the LiteLLM model string.

Optional config:

  • binary (default claude): path to the Claude Code executable.
  • system_prompt / append_system_prompt: override or extend the system prompt.
  • max_turns: cap the agentic loop iterations.
  • cwd: working directory to run claude in.
  • timeout (seconds, default 300).
  • extra_args: list of additional raw claude flags.

Note: endpoint is accepted for interface symmetry but ignored — Claude Code is local and has no endpoint URL.

handle_request​

def handle_request(request_data: Dict[str, Any]) -> Dict[str, Any]

Send a single Claude Code turn via litellm.completion.

Flow mirrors :class:ADKAgent::

request_data → litellm.completion(model="hackagent_claude_code_<id>/<model>",
                                  messages=…)
              → _ClaudeCodeCustomLLM.completion → ``claude -p``