llm-claude-cli

LLM plugin for using Claude models through the Claude Code CLI.

This plugin allows you to use Claude models via the claude command-line tool, leveraging your existing Claude Code subscription without requiring a separate API key.

Installation

First, ensure you have Claude Code CLI installed and authenticated:

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

Then install this plugin:

llm install llm-claude-cli

Or install from source:

pip install -e .

Usage

Basic usage

# Use the default Claude CLI model
llm -m claude-cli "What is the capital of France?"

# Use specific model variants
llm -m claude-cli-opus "Explain quantum computing"
llm -m claude-cli-sonnet "Write a haiku about programming"
llm -m claude-cli-haiku "Quick summary of Python decorators"

Available models

Model ID	Aliases	Description
claude-cli	ccli	Default Claude CLI model
claude-cli-opus	ccli-opus	Claude Opus 4.5
claude-cli-sonnet	ccli-sonnet	Claude Sonnet 4.5
claude-cli-haiku	ccli-haiku	Claude Haiku 4.5

Using aliases

llm -m ccli "Hello, Claude!"
llm -m ccli-opus "Deep analysis of this code"
llm -m ccli-sonnet "Explain this concept"
llm -m ccli-haiku "Quick answer"

Options

# Set timeout (in seconds, default 300)
llm -m claude-cli "Complex task" -o timeout 600

# Limit allowed tools
llm -m claude-cli "Read file.txt" -o allowedTools "Read,Glob"

# Block specific tools
llm -m claude-cli "Help me code" -o disallowedTools "Bash(rm:*),Bash(sudo:*)"

# Add additional directory for Claude to access
llm -m claude-cli "Review the frontend" -o add_dir "../frontend"

# Set working directory
llm -m claude-cli "Run the tests" -o cwd "/path/to/project"

# Set permission mode (default, acceptEdits, bypassPermissions)
llm -m claude-cli "Refactor this" -o permission_mode "acceptEdits"

# Use custom MCP configuration
llm -m claude-cli "Use my tools" -o mcp_config "/path/to/mcp.json"

# Enable verbose logging
llm -m claude-cli "Debug this" -o verbose true

System Prompts

The plugin auto-detects when to use Claude Code's default agentic system prompt:

• Simple queries (no agentic options, no schema): Uses an empty system prompt for direct responses
• Agentic queries (tools, permissions, schema, etc.): Uses Claude Code's default system prompt

You can customize this behavior:

# Simple query - uses empty system prompt by default
llm -m claude-cli "What is the capital of France?"

# Custom system prompt (replaces the default)
llm -m claude-cli "Hello" --system "You are a pirate"

# Append to Claude Code's default system prompt instead of replacing
# This preserves agentic capabilities while adding your instructions
llm -m claude-cli "Analyze this code" --system "Focus on security" -o append_system_prompt true

# Explicitly control whether to use Claude Code's default system prompt
llm -m claude-cli "Help me code" -o use_default_system_prompt true

Conversations

Use LLM's conversation features for multi-turn interactions:

# Start a conversation (gets assigned an ID automatically)
llm -m claude-cli "My name is Alice"

# Continue the most recent conversation
llm -m claude-cli "What's my name?" -c

# View conversation history and IDs
llm logs list

# Continue a specific conversation by ID
llm -m claude-cli "Tell me more" --cid 01abc123...

The plugin generates a deterministic Claude session ID from the LLM conversation ID, so continuing a conversation automatically resumes the corresponding Claude session. For one-off prompts (without -c or --cid), the plugin passes --no-session-persistence to Claude to avoid creating orphan sessions.

Streaming

Streaming is enabled by default. The plugin uses Claude Code's stream-json output format for real-time token streaming.

# Streaming output (default)
llm -m claude-cli "Tell me a story"

# Disable streaming
llm -m claude-cli "Quick answer" --no-stream

Schemas (Structured Output)

This plugin supports LLM's schema feature for structured JSON output. Use the --schema flag to specify the expected output format:

# Simple schema with field types
llm -m claude-cli "Invent a cool dog" --schema 'name, age int, breed'

# Complex schema
llm -m claude-cli "List 3 countries" --schema '
  countries: [{
    name,
    capital,
    population int
  }]
'

# Using a JSON schema file
llm -m claude-cli "Extract info" --schema-file schema.json

The plugin uses Claude Code's --json-schema flag to ensure structured output conformance.

Python API

import llm

# Get a model
model = llm.get_model("claude-cli")

# Basic prompt
response = model.prompt("What is Python?")
print(response.text())

# With system prompt
response = model.prompt("Explain decorators", system="You are a Python expert")
print(response.text())

# Streaming
for chunk in model.prompt("Tell me a story", stream=True):
    print(chunk, end="", flush=True)

# Conversation
conversation = model.conversation()
response1 = conversation.prompt("My name is Alice")
response2 = conversation.prompt("What's my name?")
print(response2.text())  # Should mention Alice

# Schema for structured output
import json

schema = {
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age": {"type": "integer"},
        "breed": {"type": "string"}
    },
    "required": ["name", "age", "breed"]
}

response = model.prompt("Invent a cool dog", schema=schema)
dog = json.loads(response.text())
print(f"Name: {dog['name']}, Age: {dog['age']}, Breed: {dog['breed']}")

How it works

This plugin invokes the Claude Code CLI (claude) as a subprocess with the -p flag for non-interactive prompts. It supports:

• Streaming: Uses --output-format stream-json for real-time NDJSON streaming
• Non-streaming: Uses --output-format json for complete responses
• Schemas: Uses --json-schema for structured JSON output conformance
• System prompts: Auto-detects when to use Claude Code's default; supports --system-prompt (replace) or --append-system-prompt
• Conversations: Integrates with LLM's -c/--cid flags; uses --no-session-persistence for one-off prompts
• Model selection: Passes --model flag to select Opus, Sonnet, or Haiku
• Directory access: Uses --add-dir for multi-directory projects
• Permission control: Uses --permission-mode and tool filtering
• MCP integration: Uses --mcp-config for custom tool servers

Requirements

• Python 3.10+
• LLM 0.26+
• Claude Code CLI installed and authenticated

License

MIT License - see LICENSE for details.