feat: new cli provider for claude code and gemini (#3083)

This commit is contained in:
Michael Neale
2025-06-27 15:55:47 +10:00
committed by GitHub
parent f5ab33cfeb
commit 0962041696
6 changed files with 972 additions and 0 deletions

View File

@@ -0,0 +1,451 @@
use anyhow::Result;
use async_trait::async_trait;
use serde_json::{json, Value};
use std::process::Stdio;
use tokio::io::{AsyncBufReadExt, BufReader};
use tokio::process::Command;
use super::base::{ConfigKey, Provider, ProviderMetadata, ProviderUsage, Usage};
use super::errors::ProviderError;
use super::utils::emit_debug_trace;
use crate::message::{Message, MessageContent};
use crate::model::ModelConfig;
use mcp_core::content::TextContent;
use mcp_core::tool::Tool;
use mcp_core::Role;
pub const CLAUDE_CODE_DEFAULT_MODEL: &str = "default";
pub const CLAUDE_CODE_KNOWN_MODELS: &[&str] = &["default"];
pub const CLAUDE_CODE_DOC_URL: &str = "https://claude.ai/cli";
#[derive(Debug, serde::Serialize)]
pub struct ClaudeCodeProvider {
command: String,
model: ModelConfig,
}
impl Default for ClaudeCodeProvider {
fn default() -> Self {
let model = ModelConfig::new(ClaudeCodeProvider::metadata().default_model);
ClaudeCodeProvider::from_env(model).expect("Failed to initialize Claude Code provider")
}
}
impl ClaudeCodeProvider {
pub fn from_env(model: ModelConfig) -> Result<Self> {
let config = crate::config::Config::global();
let command: String = config
.get_param("CLAUDE_CODE_COMMAND")
.unwrap_or_else(|_| "claude".to_string());
Ok(Self { command, model })
}
/// Filter out the Extensions section from the system prompt
fn filter_extensions_from_system_prompt(&self, system: &str) -> String {
// Find the Extensions section and remove it
if let Some(extensions_start) = system.find("# Extensions") {
// Look for the next major section that starts with #
let after_extensions = &system[extensions_start..];
if let Some(next_section_pos) = after_extensions[1..].find("\n# ") {
// Found next section, keep everything before Extensions and after the next section
let before_extensions = &system[..extensions_start];
let next_section_start = extensions_start + next_section_pos + 1;
let after_next_section = &system[next_section_start..];
format!("{}{}", before_extensions.trim_end(), after_next_section)
} else {
// No next section found, just remove everything from Extensions onward
system[..extensions_start].trim_end().to_string()
}
} else {
// No Extensions section found, return original
system.to_string()
}
}
/// Convert goose messages to the format expected by claude CLI
fn messages_to_claude_format(&self, _system: &str, messages: &[Message]) -> Result<Value> {
let mut claude_messages = Vec::new();
for message in messages {
let role = match message.role {
Role::User => "user",
Role::Assistant => "assistant",
};
let mut content_parts = Vec::new();
for content in &message.content {
match content {
MessageContent::Text(text_content) => {
content_parts.push(json!({
"type": "text",
"text": text_content.text
}));
}
MessageContent::ToolRequest(tool_request) => {
if let Ok(tool_call) = &tool_request.tool_call {
content_parts.push(json!({
"type": "tool_use",
"id": tool_request.id,
"name": tool_call.name,
"input": tool_call.arguments
}));
}
}
MessageContent::ToolResponse(tool_response) => {
if let Ok(tool_contents) = &tool_response.tool_result {
// Convert tool result contents to text
let content_text = tool_contents
.iter()
.filter_map(|content| content.as_text())
.collect::<Vec<_>>()
.join("\n");
content_parts.push(json!({
"type": "tool_result",
"tool_use_id": tool_response.id,
"content": content_text
}));
}
}
_ => {
// Skip other content types for now
}
}
}
claude_messages.push(json!({
"role": role,
"content": content_parts
}));
}
Ok(json!(claude_messages))
}
/// Parse the JSON response from claude CLI
fn parse_claude_response(
&self,
json_lines: &[String],
) -> Result<(Message, Usage), ProviderError> {
let mut all_text_content = Vec::new();
let mut usage = Usage::default();
// Join all lines and parse as a single JSON array
let full_response = json_lines.join("");
let json_array: Vec<Value> = serde_json::from_str(&full_response).map_err(|e| {
ProviderError::RequestFailed(format!("Failed to parse JSON response: {}", e))
})?;
for parsed in json_array {
if let Some(msg_type) = parsed.get("type").and_then(|t| t.as_str()) {
match msg_type {
"assistant" => {
if let Some(message) = parsed.get("message") {
// Extract text content from this assistant message
if let Some(content) = message.get("content").and_then(|c| c.as_array())
{
for item in content {
if let Some(content_type) =
item.get("type").and_then(|t| t.as_str())
{
if content_type == "text" {
if let Some(text) =
item.get("text").and_then(|t| t.as_str())
{
all_text_content.push(text.to_string());
}
}
// Skip tool_use - those are claude CLI's internal tools
}
}
}
// Extract usage information
if let Some(usage_info) = message.get("usage") {
usage.input_tokens = usage_info
.get("input_tokens")
.and_then(|v| v.as_i64())
.map(|v| v as i32);
usage.output_tokens = usage_info
.get("output_tokens")
.and_then(|v| v.as_i64())
.map(|v| v as i32);
// Calculate total if not provided
if usage.total_tokens.is_none() {
if let (Some(input), Some(output)) =
(usage.input_tokens, usage.output_tokens)
{
usage.total_tokens = Some(input + output);
}
}
}
}
}
"result" => {
// Extract additional usage info from result if available
if let Some(result_usage) = parsed.get("usage") {
if usage.input_tokens.is_none() {
usage.input_tokens = result_usage
.get("input_tokens")
.and_then(|v| v.as_i64())
.map(|v| v as i32);
}
if usage.output_tokens.is_none() {
usage.output_tokens = result_usage
.get("output_tokens")
.and_then(|v| v.as_i64())
.map(|v| v as i32);
}
}
}
_ => {} // Ignore other message types
}
}
}
// Combine all text content into a single message
let combined_text = all_text_content.join("\n\n");
if combined_text.is_empty() {
return Err(ProviderError::RequestFailed(
"No text content found in response".to_string(),
));
}
let message_content = vec![MessageContent::Text(TextContent {
text: combined_text,
annotations: None,
})];
let response_message = Message {
role: Role::Assistant,
created: chrono::Utc::now().timestamp(),
content: message_content,
};
Ok((response_message, usage))
}
async fn execute_command(
&self,
system: &str,
messages: &[Message],
_tools: &[Tool],
) -> Result<Vec<String>, ProviderError> {
let messages_json = self
.messages_to_claude_format(system, messages)
.map_err(|e| {
ProviderError::RequestFailed(format!("Failed to format messages: {}", e))
})?;
// Create a filtered system prompt without Extensions section
let filtered_system = self.filter_extensions_from_system_prompt(system);
if std::env::var("GOOSE_CLAUDE_CODE_DEBUG").is_ok() {
println!("=== CLAUDE CODE PROVIDER DEBUG ===");
println!("Command: {}", self.command);
println!("Original system prompt length: {} chars", system.len());
println!(
"Filtered system prompt length: {} chars",
filtered_system.len()
);
println!("Filtered system prompt: {}", filtered_system);
println!(
"Messages JSON: {}",
serde_json::to_string_pretty(&messages_json)
.unwrap_or_else(|_| "Failed to serialize".to_string())
);
println!("================================");
}
let mut cmd = Command::new(&self.command);
cmd.arg("-p")
.arg(messages_json.to_string())
.arg("--system-prompt")
.arg(&filtered_system)
.arg("--verbose")
.arg("--output-format")
.arg("json");
cmd.stdout(Stdio::piped()).stderr(Stdio::piped());
let mut child = cmd
.spawn()
.map_err(|e| ProviderError::RequestFailed(format!("Failed to spawn command: {}", e)))?;
let stdout = child
.stdout
.take()
.ok_or_else(|| ProviderError::RequestFailed("Failed to capture stdout".to_string()))?;
let mut reader = BufReader::new(stdout);
let mut lines = Vec::new();
let mut line = String::new();
loop {
line.clear();
match reader.read_line(&mut line).await {
Ok(0) => break, // EOF
Ok(_) => {
let trimmed = line.trim();
if !trimmed.is_empty() {
lines.push(trimmed.to_string());
}
}
Err(e) => {
return Err(ProviderError::RequestFailed(format!(
"Failed to read output: {}",
e
)));
}
}
}
let exit_status = child.wait().await.map_err(|e| {
ProviderError::RequestFailed(format!("Failed to wait for command: {}", e))
})?;
if !exit_status.success() {
return Err(ProviderError::RequestFailed(format!(
"Command failed with exit code: {:?}",
exit_status.code()
)));
}
tracing::debug!("Command executed successfully, got {} lines", lines.len());
for (i, line) in lines.iter().enumerate() {
tracing::debug!("Line {}: {}", i, line);
}
Ok(lines)
}
/// Generate a simple session description without calling subprocess
fn generate_simple_session_description(
&self,
messages: &[Message],
) -> Result<(Message, ProviderUsage), ProviderError> {
// Extract the first user message text
let description = messages
.iter()
.find(|m| m.role == mcp_core::Role::User)
.and_then(|m| {
m.content.iter().find_map(|c| match c {
MessageContent::Text(text_content) => Some(&text_content.text),
_ => None,
})
})
.map(|text| {
// Take first few words, limit to 4 words
text.split_whitespace()
.take(4)
.collect::<Vec<_>>()
.join(" ")
})
.unwrap_or_else(|| "Simple task".to_string());
if std::env::var("GOOSE_CLAUDE_CODE_DEBUG").is_ok() {
println!("=== CLAUDE CODE PROVIDER DEBUG ===");
println!("Generated simple session description: {}", description);
println!("Skipped subprocess call for session description");
println!("================================");
}
let message = Message {
role: mcp_core::Role::Assistant,
created: chrono::Utc::now().timestamp(),
content: vec![MessageContent::Text(mcp_core::content::TextContent {
text: description.clone(),
annotations: None,
})],
};
let usage = Usage::default();
Ok((
message,
ProviderUsage::new(self.model.model_name.clone(), usage),
))
}
}
#[async_trait]
impl Provider for ClaudeCodeProvider {
fn metadata() -> ProviderMetadata {
ProviderMetadata::new(
"claude-code",
"Claude Code",
"Execute Claude models via claude CLI tool",
CLAUDE_CODE_DEFAULT_MODEL,
CLAUDE_CODE_KNOWN_MODELS.to_vec(),
CLAUDE_CODE_DOC_URL,
vec![ConfigKey::new(
"CLAUDE_CODE_COMMAND",
false,
false,
Some("claude"),
)],
)
}
fn get_model_config(&self) -> ModelConfig {
// Return a custom config with 200K token limit for Claude Code
ModelConfig::new("claude-3-5-sonnet-latest".to_string()).with_context_limit(Some(200_000))
}
#[tracing::instrument(
skip(self, system, messages, tools),
fields(model_config, input, output, input_tokens, output_tokens, total_tokens)
)]
async fn complete(
&self,
system: &str,
messages: &[Message],
tools: &[Tool],
) -> Result<(Message, ProviderUsage), ProviderError> {
// Check if this is a session description request (short system prompt asking for 4 words or less)
if system.contains("four words or less") || system.contains("4 words or less") {
return self.generate_simple_session_description(messages);
}
let json_lines = self.execute_command(system, messages, tools).await?;
let (message, usage) = self.parse_claude_response(&json_lines)?;
// Create a dummy payload for debug tracing
let payload = json!({
"command": self.command,
"model": self.model.model_name,
"system": system,
"messages": messages.len()
});
let response = json!({
"lines": json_lines.len(),
"usage": usage
});
emit_debug_trace(&self.model, &payload, &response, &usage);
Ok((
message,
ProviderUsage::new(self.model.model_name.clone(), usage),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_claude_code_model_config() {
let provider = ClaudeCodeProvider::default();
let config = provider.get_model_config();
assert_eq!(config.model_name, "claude-3-5-sonnet-latest");
assert_eq!(config.context_limit(), 200_000);
}
}

View File

@@ -5,8 +5,10 @@ use super::{
azure::AzureProvider,
base::{Provider, ProviderMetadata},
bedrock::BedrockProvider,
claude_code::ClaudeCodeProvider,
databricks::DatabricksProvider,
gcpvertexai::GcpVertexAIProvider,
gemini_cli::GeminiCliProvider,
githubcopilot::GithubCopilotProvider,
google::GoogleProvider,
groq::GroqProvider,
@@ -42,8 +44,10 @@ pub fn providers() -> Vec<ProviderMetadata> {
AnthropicProvider::metadata(),
AzureProvider::metadata(),
BedrockProvider::metadata(),
ClaudeCodeProvider::metadata(),
DatabricksProvider::metadata(),
GcpVertexAIProvider::metadata(),
GeminiCliProvider::metadata(),
GithubCopilotProvider::metadata(),
GoogleProvider::metadata(),
GroqProvider::metadata(),
@@ -120,7 +124,9 @@ fn create_provider(name: &str, model: ModelConfig) -> Result<Arc<dyn Provider>>
"anthropic" => Ok(Arc::new(AnthropicProvider::from_env(model)?)),
"azure_openai" => Ok(Arc::new(AzureProvider::from_env(model)?)),
"aws_bedrock" => Ok(Arc::new(BedrockProvider::from_env(model)?)),
"claude-code" => Ok(Arc::new(ClaudeCodeProvider::from_env(model)?)),
"databricks" => Ok(Arc::new(DatabricksProvider::from_env(model)?)),
"gemini-cli" => Ok(Arc::new(GeminiCliProvider::from_env(model)?)),
"groq" => Ok(Arc::new(GroqProvider::from_env(model)?)),
"ollama" => Ok(Arc::new(OllamaProvider::from_env(model)?)),
"openrouter" => Ok(Arc::new(OpenRouterProvider::from_env(model)?)),

View File

@@ -0,0 +1,307 @@
use anyhow::Result;
use async_trait::async_trait;
use serde_json::json;
use std::process::Stdio;
use tokio::io::{AsyncBufReadExt, BufReader};
use tokio::process::Command;
use super::base::{Provider, ProviderMetadata, ProviderUsage, Usage};
use super::errors::ProviderError;
use super::utils::emit_debug_trace;
use crate::message::{Message, MessageContent};
use crate::model::ModelConfig;
use mcp_core::content::TextContent;
use mcp_core::tool::Tool;
use mcp_core::Role;
pub const GEMINI_CLI_DEFAULT_MODEL: &str = "default";
pub const GEMINI_CLI_KNOWN_MODELS: &[&str] = &["default"];
pub const GEMINI_CLI_DOC_URL: &str = "https://ai.google.dev/gemini-api/docs";
#[derive(Debug, serde::Serialize)]
pub struct GeminiCliProvider {
command: String,
model: ModelConfig,
}
impl Default for GeminiCliProvider {
fn default() -> Self {
let model = ModelConfig::new(GeminiCliProvider::metadata().default_model);
GeminiCliProvider::from_env(model).expect("Failed to initialize Gemini CLI provider")
}
}
impl GeminiCliProvider {
pub fn from_env(model: ModelConfig) -> Result<Self> {
let command = "gemini".to_string(); // Fixed command, no configuration needed
Ok(Self { command, model })
}
/// Filter out the Extensions section from the system prompt
fn filter_extensions_from_system_prompt(&self, system: &str) -> String {
// Find the Extensions section and remove it
if let Some(extensions_start) = system.find("# Extensions") {
// Look for the next major section that starts with #
let after_extensions = &system[extensions_start..];
if let Some(next_section_pos) = after_extensions[1..].find("\n# ") {
// Found next section, keep everything before Extensions and after the next section
let before_extensions = &system[..extensions_start];
let next_section_start = extensions_start + next_section_pos + 1;
let after_next_section = &system[next_section_start..];
format!("{}{}", before_extensions.trim_end(), after_next_section)
} else {
// No next section found, just remove everything from Extensions onward
system[..extensions_start].trim_end().to_string()
}
} else {
// No Extensions section found, return original
system.to_string()
}
}
/// Execute gemini CLI command with simple text prompt
async fn execute_command(
&self,
system: &str,
messages: &[Message],
_tools: &[Tool],
) -> Result<Vec<String>, ProviderError> {
// Create a simple prompt combining system + conversation
let mut full_prompt = String::new();
// Add system prompt
let filtered_system = self.filter_extensions_from_system_prompt(system);
full_prompt.push_str(&filtered_system);
full_prompt.push_str("\n\n");
// Add conversation history
for message in messages {
let role_prefix = match message.role {
Role::User => "Human: ",
Role::Assistant => "Assistant: ",
};
full_prompt.push_str(role_prefix);
for content in &message.content {
if let MessageContent::Text(text_content) = content {
full_prompt.push_str(&text_content.text);
full_prompt.push('\n');
}
}
full_prompt.push('\n');
}
full_prompt.push_str("Assistant: ");
if std::env::var("GOOSE_GEMINI_CLI_DEBUG").is_ok() {
println!("=== GEMINI CLI PROVIDER DEBUG ===");
println!("Command: {}", self.command);
println!("Full prompt: {}", full_prompt);
println!("================================");
}
let mut cmd = Command::new(&self.command);
cmd.arg("-p").arg(&full_prompt).arg("--yolo");
cmd.stdout(Stdio::piped()).stderr(Stdio::piped());
let mut child = cmd
.spawn()
.map_err(|e| ProviderError::RequestFailed(format!("Failed to spawn command: {}", e)))?;
let stdout = child
.stdout
.take()
.ok_or_else(|| ProviderError::RequestFailed("Failed to capture stdout".to_string()))?;
let mut reader = BufReader::new(stdout);
let mut lines = Vec::new();
let mut line = String::new();
loop {
line.clear();
match reader.read_line(&mut line).await {
Ok(0) => break, // EOF
Ok(_) => {
let trimmed = line.trim();
if !trimmed.is_empty() {
lines.push(trimmed.to_string());
}
}
Err(e) => {
return Err(ProviderError::RequestFailed(format!(
"Failed to read output: {}",
e
)));
}
}
}
let exit_status = child.wait().await.map_err(|e| {
ProviderError::RequestFailed(format!("Failed to wait for command: {}", e))
})?;
if !exit_status.success() {
return Err(ProviderError::RequestFailed(format!(
"Command failed with exit code: {:?}",
exit_status.code()
)));
}
tracing::debug!(
"Gemini CLI executed successfully, got {} lines",
lines.len()
);
Ok(lines)
}
/// Parse simple text response
fn parse_response(&self, lines: &[String]) -> Result<(Message, Usage), ProviderError> {
// Join all lines into a single response
let response_text = lines.join("\n");
if response_text.trim().is_empty() {
return Err(ProviderError::RequestFailed(
"Empty response from gemini command".to_string(),
));
}
let message = Message {
role: Role::Assistant,
created: chrono::Utc::now().timestamp(),
content: vec![MessageContent::Text(TextContent {
text: response_text,
annotations: None,
})],
};
let usage = Usage::default(); // No usage info available for gemini CLI
Ok((message, usage))
}
/// Generate a simple session description without calling subprocess
fn generate_simple_session_description(
&self,
messages: &[Message],
) -> Result<(Message, ProviderUsage), ProviderError> {
// Extract the first user message text
let description = messages
.iter()
.find(|m| m.role == Role::User)
.and_then(|m| {
m.content.iter().find_map(|c| match c {
MessageContent::Text(text_content) => Some(&text_content.text),
_ => None,
})
})
.map(|text| {
// Take first few words, limit to 4 words
text.split_whitespace()
.take(4)
.collect::<Vec<_>>()
.join(" ")
})
.unwrap_or_else(|| "Simple task".to_string());
if std::env::var("GOOSE_GEMINI_CLI_DEBUG").is_ok() {
println!("=== GEMINI CLI PROVIDER DEBUG ===");
println!("Generated simple session description: {}", description);
println!("Skipped subprocess call for session description");
println!("================================");
}
let message = Message {
role: Role::Assistant,
created: chrono::Utc::now().timestamp(),
content: vec![MessageContent::Text(TextContent {
text: description.clone(),
annotations: None,
})],
};
let usage = Usage::default();
Ok((
message,
ProviderUsage::new(self.model.model_name.clone(), usage),
))
}
}
#[async_trait]
impl Provider for GeminiCliProvider {
fn metadata() -> ProviderMetadata {
ProviderMetadata::new(
"gemini-cli",
"Gemini CLI",
"Execute Gemini models via gemini CLI tool",
GEMINI_CLI_DEFAULT_MODEL,
GEMINI_CLI_KNOWN_MODELS.to_vec(),
GEMINI_CLI_DOC_URL,
vec![], // No configuration needed
)
}
fn get_model_config(&self) -> ModelConfig {
// Return a custom config with 1M token limit for Gemini CLI
ModelConfig::new("gemini-1.5-pro".to_string()).with_context_limit(Some(1_000_000))
}
#[tracing::instrument(
skip(self, system, messages, tools),
fields(model_config, input, output, input_tokens, output_tokens, total_tokens)
)]
async fn complete(
&self,
system: &str,
messages: &[Message],
tools: &[Tool],
) -> Result<(Message, ProviderUsage), ProviderError> {
// Check if this is a session description request (short system prompt asking for 4 words or less)
if system.contains("four words or less") || system.contains("4 words or less") {
return self.generate_simple_session_description(messages);
}
let lines = self.execute_command(system, messages, tools).await?;
let (message, usage) = self.parse_response(&lines)?;
// Create a dummy payload for debug tracing
let payload = json!({
"command": self.command,
"model": self.model.model_name,
"system": system,
"messages": messages.len()
});
let response = json!({
"lines": lines.len(),
"usage": usage
});
emit_debug_trace(&self.model, &payload, &response, &usage);
Ok((
message,
ProviderUsage::new(self.model.model_name.clone(), usage),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_gemini_cli_model_config() {
let provider = GeminiCliProvider::default();
let config = provider.get_model_config();
assert_eq!(config.model_name, "gemini-1.5-pro");
assert_eq!(config.context_limit(), 1_000_000);
}
}

View File

@@ -3,6 +3,7 @@ pub mod azure;
pub mod azureauth;
pub mod base;
pub mod bedrock;
pub mod claude_code;
pub mod databricks;
pub mod embedding;
pub mod errors;
@@ -10,6 +11,7 @@ mod factory;
pub mod formats;
mod gcpauth;
pub mod gcpvertexai;
pub mod gemini_cli;
pub mod githubcopilot;
pub mod google;
pub mod groq;

View File

@@ -36,6 +36,19 @@ Goose relies heavily on tool calling capabilities and currently works best with
| [Venice AI](https://venice.ai/home) | Provides access to open source models like Llama, Mistral, and Qwen while prioritizing user privacy. **Requires an account and an [API key](https://docs.venice.ai/overview/guides/generating-api-key)**. | `VENICE_API_KEY`, `VENICE_HOST` (optional), `VENICE_BASE_PATH` (optional), `VENICE_MODELS_PATH` (optional) |
| [xAI](https://x.ai/) | Access to xAI's Grok models including grok-3, grok-3-mini, and grok-3-fast with 131,072 token context window. | `XAI_API_KEY`, `XAI_HOST` (optional) |
## CLI Providers
Goose also supports special "pass-through" providers that work with existing CLI tools, allowing you to use your subscriptions instead of paying per token:
| Provider | Description | Requirements |
|-----------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [Claude Code](https://claude.ai/cli) (`claude-code`) | Uses Anthropic's Claude CLI tool with your Claude Code subscription. Provides access to Claude with 200K context limit. | Claude CLI installed and authenticated, active Claude Code subscription |
| [Gemini CLI](https://ai.google.dev/gemini-api/docs) (`gemini-cli`) | Uses Google's Gemini CLI tool with your Google AI subscription. Provides access to Gemini with 1M context limit. | Gemini CLI installed and authenticated |
:::tip CLI Providers
CLI providers are cost-effective alternatives that use your existing subscriptions. They work differently from API providers as they execute CLI commands and integrate with the tools' native capabilities. See the [CLI Providers guide](/docs/guides/cli-providers) for detailed setup instructions.
:::
## Configure Provider

View File

@@ -0,0 +1,193 @@
---
sidebar_position: 25
title: CLI Providers
sidebar_label: CLI Providers
---
# CLI Providers
Goose supports two special "pass-through" providers that leverage existing CLI tools from Anthropic and Google. These providers allow you to use your existing subscriptions to Claude Code and Google Gemini CLI tools directly through Goose, providing a cost-effective way to access these excellent AI models and tools while maintaining the benefits of Goose's ecosystem, and working with multiple models and extensions.
## Overview
CLI providers are different from traditional LLM providers in several key ways:
- **Pass-through architecture**: Instead of making direct API calls, these providers drive the CLI commands
- **Subscription-based**: Use your existing Claude Code or Google Gemini CLI subscriptions instead of paying per token (usually a flat monthly fee)
- **Tool integration**: These providers work with their respective CLI tools' built-in capabilities while Goose manages sessions and extensions
- **Cost-effective**: Leverage unlimited or subscription-based pricing models instead of token-based billing
## Available CLI Providers
### Claude Code (`claude-code`)
The Claude Code provider integrates with Anthropic's [Claude CLI tool](https://claude.ai/cli), allowing you to use Claude models through your existing Claude Code subscription.
**Features:**
- Uses Claude's latest models
- 200,000 token context limit
- Automatic filtering of Goose extensions from system prompts (since Claude Code has its own tool ecosystem)
- JSON output parsing for structured responses
**Requirements:**
- Claude CLI tool installed and configured
- Active Claude Code subscription
- CLI tool authenticated with your Anthropic account
### Gemini CLI (`gemini-cli`)
The Gemini CLI provider integrates with Google's [Gemini CLI tool](https://ai.google.dev/gemini-api/docs), providing access to Gemini models through your Google AI subscription.
**Features:**
- 1,000,000 token context limit
**Requirements:**
- Gemini CLI tool installed and configured
- CLI tool authenticated with your Google account
## Setup Instructions
### Setting up Claude Code Provider
1. **Install Claude CLI Tool**
Follow the installation instructions [here](https://docs.anthropic.com/en/docs/claude-code/overview) to install and configure the Claude CLI tool.
2. **Authenticate with Claude**
Ensure your Claude CLI is authenticated and working
3. **Configure Goose**
Set the provider environment variable:
```bash
export GOOSE_PROVIDER=claude-code
```
Or configure through the Goose CLI:
```bash
goose configure
# Select "Configure Providers"
# Choose "Claude Code" from the list
```
### Setting up Gemini CLI Provider
1. **Install Gemini CLI Tool**
Follow the installation instructions for [gemini cli]https://blog.google/technology/developers/introducing-gemini-cli-open-source-ai-agent/) to install and configure the Gemini CLI tool.
2. **Authenticate with Google**
Ensure your Gemini CLI is authenticated and working.
3. **Configure Goose**
Set the provider environment variable:
```bash
export GOOSE_PROVIDER=gemini-cli
```
Or configure through the Goose CLI:
```bash
goose configure
# Select "Configure Providers"
# Choose "Gemini CLI" from the list
```
## Usage Examples
### Basic Usage
Once configured, use these providers just like any other Goose provider:
```bash
# Using Claude Code
GOOSE_PROVIDER=claude-code goose session start
# Using Gemini CLI
GOOSE_PROVIDER=gemini-cli goose session start
```
### Combining with Other Models
CLI providers work well in combination with other models using Goose's lead/worker pattern:
```bash
# Use Claude Code as lead model, GPT-4o as worker
export GOOSE_LEAD_PROVIDER=claude-code
export GOOSE_PROVIDER=openai
export GOOSE_MODEL=gpt-4o
export GOOSE_LEAD_MODEL=default
goose session start
```
## Configuration Options
### Claude Code Configuration
| Environment Variable | Description | Default |
|---------------------|-------------|---------|
| `GOOSE_PROVIDER` | Set to `claude-code` to use this provider | None |
| `CLAUDE_CODE_COMMAND` | Path to the Claude CLI command | `claude` |
### Gemini CLI Configuration
| Environment Variable | Description | Default |
|---------------------|-------------|---------|
| `GOOSE_PROVIDER` | Set to `gemini-cli` to use this provider | None |
## How It Works
### System Prompt Filtering
Both CLI providers automatically filter out Goose's extension information from system prompts since these CLI tools have their own tool ecosystems. This prevents conflicts and ensures clean interaction with the underlying CLI tools.
### Message Translation
- **Claude Code**: Converts Goose messages to Claude's JSON message format, handling tool calls and responses appropriately
- **Gemini CLI**: Converts messages to simple text prompts with role prefixes (Human:/Assistant:)
### Response Processing
- **Claude Code**: Parses JSON responses to extract text content and usage information
- **Gemini CLI**: Processes plain text responses from the CLI tool
## Limitations and Considerations
### Tool Calling
These providers handle tool calling differently than standard API providers as they have their own tool integrations out of the box.
### Error Handling
CLI providers depend on external tools, so ensure:
- CLI tools are properly installed and in your PATH
- Authentication is maintained and valid
- Subscription limits are not exceeded
## Benefits
### Cost Effectiveness
- **Subscription-based pricing**: Use unlimited or fixed-price subscriptions instead of per-token billing
- **Existing subscriptions**: Leverage subscriptions you may already have
### Flexibility
- **Hybrid usage**: Combine CLI providers with API-based providers using lead/worker patterns
- **Session management**: Full Goose session management and extension system
- **Easy switching**: Switch between CLI and API providers as needed
## Best Practices
1. **Authentication Management**: Keep CLI tools authenticated and monitor subscription status
2. **Hybrid Approaches**: Consider using CLI providers for heavy workloads and API providers for quick tasks
3. **Backup Providers**: Configure fallback API providers in case CLI tools are unavailable
---
CLI providers offer a powerful way to integrate Goose with existing AI tool subscriptions while maintaining the benefits of Goose's session management and extension ecosystem. They're particularly valuable for users with existing subscriptions who want to maximize their investment while gaining access to Goose's capabilities.