ccrust

The claude-code-rust repository builds the ccrust binary: a Rust-native reimplementation of Claude Code agent workflows with support for tasks, tools, plugins, MCP, and multiple providers, including Anthropic, native Gemini, ChatGPT Codex, and OpenAI-compatible backends such as OpenRouter. It is targeted fully compatible with Claude's official Claude Code with fully rewriting with Rust.

You can use your own Codex Plan with Claude Code's agent workflow.

We also introduced some prompt compacting but strictly keeping Claude Code's original prompt caching and format, and fixed some bugs of token cost in original Claude Code. It should work exactly the same way with original Claude Code.

In ZanChat AI, we use this tool with Codex Plan for production-level development. As we are updating the tools everyday, so we recommend you to upgrade it frequently. We'll try our best to keep it stable.

Installation

Requires Rust and Cargo.

git clone git@github.com:ZanChat/claude-code-rust.git
cd claude-code-rust

cargo install --path crates/cli --locked

Cargo installs the binary as ccrust into ~/.cargo/bin by default.

To refresh a local build while testing changes:

cargo install --path crates/cli --locked --force

Running ccrust

# Check the installed binary
ccrust --version

# Interactive REPL
ccrust

# Non-interactive prompt
ccrust 'Refactor the auth logic in src/auth.rs'

# Headless JSON output
ccrust --print --output-format json 'Summarize the current diff'

# Pick a provider explicitly for one launch
ccrust --provider chatgpt-codex

ccrust loads saved login config from ~/.claude/.env.ccrust and then ./.claude/.env.ccrust. Later sources override earlier ones: project-local config overrides user config, real environment variables override saved config, and CLI flags override all of them.

The top-level CLI now exposes real backends for these command families:

• ccrust mcp, ccrust plugin, ccrust agents, ccrust doctor, ccrust setup-token
• ccrust server and headless ccrust open <endpoint> -p [prompt]
• ccrust update, ccrust install, ccrust rollback
• ccrust task, ccrust export, ccrust completion

Current compatibility gap: TS direct-connect cc:// and cc+unix:// URLs still are not implemented in the Rust rewrite, and ccrust ssh still reports that remote SSH session orchestration is not available yet.

If no provider is configured, bare ccrust opens onboarding automatically. Inside the REPL, /login opens the same onboarding flow and /logout clears the tracked *.env.ccrust config for that session.

For OpenAI-family auth compatibility, ccrust uses the same shared Codex credentials file as the TS version and official codex login: auth.json is resolved from CODEX_AUTH_JSON_PATH, then CLAUDE_CODE_CODEX_AUTH_JSON_PATH, then CODEX_HOME / CLAUDE_CODE_CODEX_HOME, and finally ~/.codex/auth.json.

Any ccrust-only fallback auth snapshot is stored separately at ~/.claude/ccrust/code-agent-auth.json by default, so using ccrust does not move or replace the shared auth.json that other tools continue to use.

Supported API Providers & Authentication

1. Anthropic (First-Party) — Default

export ANTHROPIC_API_KEY="sk-ant-api..."
ccrust

2. OpenAI-Compatible

For OpenAI-family providers, the agent uses a dual-model architecture matching the original TS implementation:

• reasoning model: REASONING_MODEL (default gpt-5.4) for thinking-enabled turns
• completion model: COMPLETION_MODEL (default gpt-5.3-codex) for standard turns and utility calls

The agent automatically selects the appropriate model per request based on whether reasoning is active for a given turn.

export OPENAI_API_KEY="sk-..."
ccrust --provider openai-compatible

Official OpenAI uses the same openai-compatible provider mode and defaults to https://api.openai.com/v1 when OPENAI_BASE_URL is unset. The legacy openai provider name is still accepted as an alias for compatibility.

For other OpenAI-compatible APIs:

export OPENAI_API_KEY="your-custom-token"
export OPENAI_BASE_URL="https://api.yourprovider.com/v1"
ccrust --provider openai-compatible

The interactive /login flow also includes presets for OpenAI, OpenRouter, Gemini, and a custom OpenAI-compatible base URL.

3. Gemini — Native Gemini API

Native Gemini is the preferred way to use Gemini with ccrust.

export GEMINI_API_KEY="..."
ccrust --provider gemini

The native Gemini provider defaults to:

• reasoning model: GEMINI_REASONING_MODEL or gemini-2.5-pro
• fast model: GEMINI_COMPLETION_MODEL or gemini-2.5-flash
• base URL: GEMINI_BASE_URL or https://generativelanguage.googleapis.com/v1beta

GOOGLE_API_KEY and OPENAI_API_KEY are also accepted. For model/base URL compatibility, native Gemini also falls back to REASONING_MODEL, COMPLETION_MODEL, REASONING_MODEL_THINK, COMPLETION_MODEL_THINK, and OPENAI_BASE_URL. If OPENAI_BASE_URL points at Gemini's OpenAI-compatible /openai endpoint, ccrust normalizes it back to the native generateContent base URL automatically. The interactive /login flow still saves provider-specific Gemini keys by default.

4. ChatGPT Codex

Uses the shared Codex auth.json for authentication with automatic token refresh. Path resolution matches the TS version and official Codex tooling: CODEX_AUTH_JSON_PATH -> CLAUDE_CODE_CODEX_AUTH_JSON_PATH -> CODEX_HOME / CLAUDE_CODE_CODEX_HOME -> ~/.codex/auth.json.

ccrust --provider chatgpt-codex

/logout does not delete the shared Codex auth.json; it only clears ccrust-managed local config and fallback snapshot state.

5. Amazon Bedrock

export AWS_ACCESS_KEY_ID="..."
export AWS_SECRET_ACCESS_KEY="..."
export AWS_REGION="us-east-1"
ccrust --provider bedrock

6. Google Cloud Vertex AI

export VERTEX_ACCESS_TOKEN="..."
ccrust --provider vertex

7. Azure AI Foundry

export ANTHROPIC_FOUNDRY_API_KEY="..."
ccrust --provider foundry

Environment Variables Reference

Provider Selection

Variable	Description	Default
CLAUDE_CODE_API_PROVIDER	Override the active provider (firstParty, gemini, openai-compatible, chatgpt-codex, bedrock, vertex, foundry). openai is accepted as a legacy alias for openai-compatible.	firstParty

Authentication

Variable	Description
ANTHROPIC_API_KEY	API key for Anthropic first-party provider
GEMINI_API_KEY	API key for the native Gemini provider (GOOGLE_API_KEY and OPENAI_API_KEY also work)
GOOGLE_API_KEY	Alternative API key for the native Gemini provider
OPENAI_API_KEY	API key / bearer token for OpenAI-family providers
AWS_ACCESS_KEY_ID	AWS access key for Bedrock
AWS_SECRET_ACCESS_KEY	AWS secret key for Bedrock
AWS_SESSION_TOKEN	Optional AWS session token for Bedrock
AWS_BEARER_TOKEN_BEDROCK	Direct bearer token for Bedrock
VERTEX_ACCESS_TOKEN	OAuth access token for Vertex AI
GOOGLE_OAUTH_ACCESS_TOKEN	Alternative Google OAuth token for Vertex AI
ANTHROPIC_FOUNDRY_API_KEY	API key for Azure AI Foundry
AZURE_API_KEY	Alternative API key for Azure AI Foundry
AZURE_AUTH_TOKEN	Bearer token for Azure AI Foundry
FOUNDRY_AUTH_TOKEN	Alternative bearer token for Foundry
CODEX_AUTH_JSON_PATH	Explicit path to the shared Codex/OpenAI auth file
CLAUDE_CODE_CODEX_AUTH_JSON_PATH	Alternate explicit path to the shared Codex/OpenAI auth file
CODEX_HOME	Base directory for Codex state such as auth.json
CLAUDE_CODE_CODEX_HOME	Alternate base directory for Codex state such as auth.json

Base URLs

Variable	Description	Default
ANTHROPIC_BASE_URL	Override Anthropic API endpoint	https://api.anthropic.com
GEMINI_BASE_URL	Override Gemini API endpoint (OPENAI_BASE_URL also works and /openai is normalized)	https://generativelanguage.googleapis.com/v1beta
OPENAI_BASE_URL	Override OpenAI API endpoint	https://api.openai.com/v1
ANTHROPIC_BEDROCK_BASE_URL / BEDROCK_BASE_URL	Override Bedrock endpoint	Auto-detected from region
ANTHROPIC_VERTEX_BASE_URL / VERTEX_BASE_URL	Override Vertex AI endpoint	Auto-detected from project/region
ANTHROPIC_FOUNDRY_BASE_URL / FOUNDRY_BASE_URL	Override Foundry endpoint	Derived from resource name
ANTHROPIC_FOUNDRY_RESOURCE	Azure AI Foundry resource name	—

Model Selection

Variable	Description	Default
REASONING_MODEL	Model for thinking-enabled turns	gpt-5.4
COMPLETION_MODEL	Model for standard and utility turns	gpt-5.3-codex
GEMINI_REASONING_MODEL	Default model for native Gemini main turns (REASONING_MODEL also works)	gemini-2.5-pro
GEMINI_COMPLETION_MODEL	Default model for native Gemini fast turns (COMPLETION_MODEL also works)	gemini-2.5-flash

OpenAI Thinking

Variable	Description	Default
REASONING_MODEL_THINK	Reasoning effort for the reasoning model (low, medium, high, xhigh); native Gemini also uses this when CLAUDE_CODE_API_PROVIDER=gemini	xhigh
COMPLETION_MODEL_THINK	Reasoning effort for the completion model (low, medium, high, xhigh); native Gemini also uses this when CLAUDE_CODE_API_PROVIDER=gemini	xhigh

Claude Thinking

Variable	Description	Default
CLAUDE_CODE_DISABLE_THINKING	Disable thinking entirely for Claude models	false
CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING	Force budget-based thinking instead of adaptive	false
MAX_THINKING_TOKENS	Token budget for non-adaptive thinking	10000

Region & Project

Variable	Description	Default
AWS_REGION / BEDROCK_AWS_REGION / AWS_DEFAULT_REGION	AWS region for Bedrock	us-east-1
CLAUDE_CODE_VERTEX_REGION / CLOUD_ML_REGION	GCP region for Vertex AI	us-east5
CLAUDE_CODE_VERTEX_PROJECT_ID / GOOGLE_CLOUD_PROJECT	GCP project for Vertex AI	—

Auth Skip Flags

Variable	Description
CLAUDE_CODE_SKIP_BEDROCK_AUTH	Skip AWS SigV4 auth for Bedrock
CLAUDE_CODE_SKIP_VERTEX_AUTH	Skip OAuth for Vertex AI
CLAUDE_CODE_SKIP_FOUNDRY_AUTH	Skip auth for Foundry

Retry

Variable	Description	Default
LLM_RETRY_COUNT	Max retries on 50x errors	3

Built-In Slash Commands

Inside the REPL:

Command	Description
/vim	Toggle Vim mode
/status	Print runtime provider and environment status
/ide	Inspect IDE bridge compatibility and connection state
/model <name>	Switch active model
/files	View context files
/diff	View context diffs
/clear	Reset conversation
/compact	Compact context to save tokens
/help	Show all available commands including plugins

Top-Level Command Notes

Print Mode

ccrust --print runs the root command in non-interactive mode. The Rust CLI supports the TS-style print contract for:

• --output-format text|json|stream-json
• --input-format text|stream-json
• --json-schema <schema>
• --include-hook-events
• --include-partial-messages
• --replay-user-messages
• --resume-session-at <assistant-message-id>
• --no-session-persistence

Local Install Management

ccrust update and ccrust install [target] install the current binary, a local executable path, or a previously recorded snapshot into ~/.cargo/bin/ccrust. Each install records rollback snapshots under ~/.claude/ccrust/install-history/ and stores state in ~/.claude/ccrust/install-state.json.

ccrust rollback [target] restores one of those recorded snapshots. Use ccrust rollback --list to inspect available entries and ccrust rollback --dry-run to preview the selected target.

Tasks, Export, and Completion

ccrust task supports:

• task create <subject> [-d|--description <text>] [-l|--list <id>]
• task list [--pending] [--json] [-l|--list <id>]
• task get <id> [-l|--list <id>]
• task update <id> [--subject <text>] [-d|--description <text>] [-s|--status <status>] [--owner <agentId>] [--clear-owner] [-l|--list <id>]
• task dir [-l|--list <id>]

ccrust export <source> <outputFile> renders a transcript to text from a session ID, a recent session index, or an existing .json / .jsonl transcript path.

ccrust completion <shell> [--output <file>] generates shell completion scripts for bash, zsh, and fish.

License

MIT