Open-source universal agent harness — Claude Code-like capabilities for any AI provider.

OpenClaudia is a Rust-based CLI that transforms any LLM into an agentic coding assistant with tools, memory, hooks, and multi-provider support.

• Behavioral Modes — Three-axis model (agency, quality, scope) with 8 presets and 6 modifiers for fine-grained control over AI behavior
• Multi-Provider Support — Anthropic, OpenAI, Google Gemini, DeepSeek, Qwen, Z.AI/GLM, Kimi/Moonshot, MiniMax, Ollama, and any OpenAI-compatible server
• Local LLM Support — Run with Ollama, LM Studio, LocalAI, or any OpenAI-compatible endpoint
• Auto-Detect Provider — Pass -m gemini-3.5-flash and the provider is detected automatically
• 30+ Agentic Tools — Bash, file ops, LSP, web search, notebooks, task tracking, plan mode, worktrees, cron scheduling, MCP resources
• Tool Execution Loop — Multi-turn tool calling with automatic result feedback (works across all providers)
• Web Search — Free DuckDuckGo/Bing browser scraping in default builds; no search API key required
• Auto-Learning Memory — Automatically captures coding patterns, error resolutions, file relationships, and user preferences across sessions
• Background Shells — Run long-running processes, check output, and kill them on demand
• Thinking Mode — Extended reasoning for Anthropic, OpenAI GPT-5/o1/o3/o4, Gemini 3.x/2.5, DeepSeek V4, Qwen QwQ, Z.AI/GLM, and MiniMax-M3
• VDD Adversarial Review — Verification-Driven Development: a separate adversary model reviews code for bugs/vulnerabilities
• Hooks System — Run custom scripts at key moments (session start, tool use, prompt submit, etc.)
• Guardrails — Configurable code quality gates, blast radius limiting, and diff size monitoring
• Plan Mode — Toggle between Build and Plan modes; plan mode restricts destructive tools
• Permissions — Granular tool-level allow/deny rules with glob patterns
• Task Management — Built-in task tracking with dependencies and status workflow
• LSP Integration — Language Server Protocol support for go-to-definition, find-references, hover, and more
• Subagent System — Spawn autonomous agents from the agent loop; coordinator infrastructure is experimental and not wired into the default TUI yet
• ACP Server — Agent Control Protocol server for agent interoperability via stdin/stdout
• Git Worktrees — Create, list, and safely remove isolated git worktrees without mutating the process CWD
• Cron Scheduling — Create, list, and delete cron schedule metadata for external schedulers
• Skills System — Load and invoke reusable prompt skills from markdown files
• Cross-Platform — Windows, macOS, Linux with Git Bash for consistent shell behavior
• Interactive TUI — Rich terminal interface with keybindings, themes, and session management
• Context Compaction — Automatic summarization when conversations get long
• Notebook Support — Read and edit Jupyter notebooks
• MCP Integration — Browse and read resources from MCP servers
• Plugin System — Install, manage, and extend with plugins (commands, hooks, MCP servers)
• OAuth Support — Use your Claude Max subscription via built-in OAuth proxy

• Rust — Install via rustup
• Git Bash (Windows only) — Comes with Git for Windows
  • OpenClaudia uses Git Bash on Windows for Unix command compatibility
  • Ensure Git is in your PATH

# Clone the repository
git clone https://github.com/dollspace-gay/openclaudia.git
cd openclaudia

# Build release version (includes browser/web search support by default)
cargo build --release

# Build without browser feature (lighter binary, no headless Chrome)
cargo build --release --no-default-features

# The binary is at target/release/openclaudia

# Set your API key (choose your provider)
export ANTHROPIC_API_KEY="your-key-here"
# or: export OPENAI_API_KEY="your-key-here"
# or: export GOOGLE_API_KEY="your-key-here"
# or: export DEEPSEEK_API_KEY="your-key-here"

# Initialize configuration in your project
openclaudia init

# Start chatting (uses default provider from config)
openclaudia

# Use a specific model (provider auto-detected from model name)
openclaudia -m gemini-3.5-flash
openclaudia -m gpt-5.5
openclaudia -m claude-sonnet-4-6

# Start with a behavioral mode
openclaudia --mode create     # Autonomous architect — build from scratch
openclaudia --mode safe       # Collaborative minimal — surgical precision
openclaudia --mode debug      # Investigation-first debugging

Configuration is stored in .openclaudia/config.yaml:

proxy:
  port: 8080
  host: "127.0.0.1"
  # Provider: anthropic, openai, google/gemini, deepseek, qwen/alibaba,
  # zai/glm/zhipu, kimi/moonshot, minimax, ollama, local, lmstudio,
  # localai, text-generation-webui
  target: anthropic

providers:
  anthropic:
    base_url: https://api.anthropic.com
    thinking:
      enabled: false
      reasoning_effort: "high"    # Claude Opus 4.8/4.7/Fable/Mythos adaptive effort
      # budget_tokens: 10000      # Manual-thinking Claude models only
  openai:
    base_url: https://api.openai.com
    thinking:
      reasoning_effort: "medium"  # OpenAI GPT-5/o1/o3/o4: none, low, medium, high, xhigh
  google:
    base_url: https://generativelanguage.googleapis.com
    thinking:
      budget_tokens: 10000        # Google Gemini thinking budget
  zai:
    base_url: https://api.z.ai/api/coding/paas/v4
  deepseek:
    base_url: https://api.deepseek.com
  qwen:
    base_url: https://dashscope.aliyuncs.com/compatible-mode
  kimi:
    base_url: https://api.moonshot.ai/v1
  minimax:
    base_url: https://api.minimax.io/v1
  # Ollama for local LLM inference
  ollama:
    base_url: http://localhost:11434
  # Any OpenAI-compatible local server (LM Studio, LocalAI, text-generation-webui, etc.)
  local:
    base_url: http://localhost:1234/v1
  lmstudio:
    base_url: http://localhost:1234/v1
  localai:
    base_url: http://localhost:8080/v1
  text-generation-webui:
    base_url: http://localhost:5000/v1

session:
  timeout_minutes: 30
  persist_path: .openclaudia/session
  max_turns: 0  # 0 = unlimited agentic loop iterations; set nonzero to cap tool loops

# Verification-Driven Development (VDD) - Adversarial code review
# vdd:
#   enabled: true
#   mode: advisory           # advisory (single pass) or blocking (loop until clean)
#   adversary:
#     provider: google       # Must differ from proxy.target
#     model: gemini-3.5-flash

# Granular tool permissions
# permissions:
#   enabled: true
#   default_allow:
#     - "git status"
#     - "src/**"
#   mcp:
#     filesystem: ["read_file", "list_directory"]

# Legacy line REPL keybindings (`openclaudia --tui-mode`)
keybindings:
  ctrl-x n: new_session
  ctrl-x x: export
  tab: toggle_mode
  escape: cancel

openclaudia                    # Start full-screen interactive TUI (default)
openclaudia -m <model>         # Use specific model (auto-detects provider)
openclaudia -v                 # Verbose logging
openclaudia --resume           # Resume last session
openclaudia --session-id <id>  # Resume specific session
openclaudia --coordinator --tui-mode  # Legacy REPL coordinator prompt mode
openclaudia --tui-mode         # Legacy line-oriented REPL
openclaudia --mode <preset>    # Start with a behavioral mode preset
openclaudia --print "prompt"   # Send one prompt, print the response, and exit

openclaudia init               # Initialize config in current directory
openclaudia init --force       # Overwrite existing config

openclaudia auth               # Authenticate with Claude Max (OAuth)
openclaudia auth --status      # Check auth status
openclaudia auth --logout      # Clear native OAuth session cache

openclaudia start              # Start as proxy server
openclaudia start -p 9090      # Custom port
openclaudia start -t openai    # Target specific provider

openclaudia acp                # Start ACP server on stdin/stdout
openclaudia acp -m <model>     # ACP with specific model

openclaudia loop               # Start iteration mode with Stop hooks
openclaudia loop -n 10         # Max 10 iterations

openclaudia config             # Show current configuration
openclaudia doctor             # Check connectivity and API keys

The default full-screen TUI intentionally exposes a focused slash-command set. The legacy line-oriented REPL (openclaudia --tui-mode) has additional commands; type /help there for that registry.

The keybindings: config map customizes the legacy line-oriented REPL (openclaudia --tui-mode). The default full-screen TUI currently uses the shortcuts above.

The lists below are the built-in /model list fallback catalog. Model names are not limited to this catalog: openclaudia -m <model> and /model <model> accept any upstream chat model ID that the selected provider endpoint supports.

• claude-fable-5, claude-mythos-5, claude-mythos-preview — Latest/highest-capability Claude 5 family
• claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6 — Claude 4 family
• claude-haiku-4-5-20251001, claude-haiku-4-5 — Fast, near-frontier
• claude-sonnet-4-5-20250929, claude-sonnet-4-5, claude-opus-4-5-20251101, claude-opus-4-5, claude-opus-4-1-20250805 — Legacy

• gpt-5.5, gpt-5.5-pro, gpt-5.5-2026-04-23, gpt-5.5-pro-2026-04-23 — Latest frontier family
• gpt-5.4, gpt-5.4-pro, gpt-5.4-2026-03-05, gpt-5.4-pro-2026-03-05, gpt-5.4-mini, gpt-5.4-mini-2026-03-17, gpt-5.4-nano, gpt-5.4-nano-2026-03-17 — Current GPT-5.4 family
• gpt-5.3-codex, gpt-5.3-chat-latest, gpt-5.2, gpt-5.2-pro, gpt-5.2-2025-12-11, gpt-5.2-pro-2025-12-11, gpt-5.2-codex, gpt-5.2-chat-latest — Codex/previous frontier family
• gpt-5.1, gpt-5.1-2025-11-13, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.1-chat-latest — GPT-5.1 family
• gpt-5, gpt-5-pro, gpt-5-2025-08-07, gpt-5-pro-2025-10-06, gpt-5-codex, gpt-5-chat-latest, gpt-5-chat-latest-2025-08-07, gpt-5-mini, gpt-5-mini-2025-08-07, gpt-5-nano, gpt-5-nano-2025-08-07 — GPT-5 family
• gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, gpt-4.1-nano-2025-04-14, o3-pro, o3-pro-2025-06-10, o3, o3-2025-04-16, o3-mini, o3-mini-2025-01-31, o4-mini, o4-mini-2025-04-16, o1-pro, o1-pro-2025-03-19, o1, o1-2024-12-17, o1-mini, o1-mini-2024-09-12, o1-preview — Legacy chat/reasoning models
• chat-latest, gpt-4o-search-preview, gpt-4o-mini, gpt-4o-mini-2024-07-18, gpt-4o-mini-search-preview, gpt-4o, gpt-4o-2024-11-20, gpt-4o-2024-08-06, gpt-4.5-preview, gpt-4-turbo, gpt-4-turbo-2024-04-09, gpt-4-turbo-preview, gpt-4, gpt-4-0613, gpt-3.5-turbo, gpt-3.5-turbo-0125, codex-mini-latest — Compatibility and deprecated chat models

• gemini-3.5-flash, gemini-3.1-pro-preview, gemini-3.1-pro-preview-customtools, gemini-3.1-flash-lite, gemini-3-flash-preview — Gemini 3 family
• gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite — Stable GA

• deepseek-v4-pro, deepseek-v4-flash — DeepSeek V4 family
• deepseek-chat, deepseek-reasoner — Compatibility aliases for V4 Flash through 2026-07-24

• qwen3.7-max, qwen3.7-max-2026-06-08, qwen3.7-max-2026-05-20, qwen3.7-max-2026-05-17, qwen3.7-max-preview, qwen3.6-max-preview, qwen3-max, qwen3-max-2026-01-23, qwen3-max-2025-09-23, qwen3-max-preview, qwen-max — Qwen Max family
• qwen3.7-plus, qwen3.7-plus-2026-05-26, qwen3.6-plus, qwen3.6-plus-2026-04-02, qwen3.5-plus, qwen3.5-plus-2026-04-20, qwen3.5-plus-2026-02-15, qwen-plus, qwen-plus-latest, qwen-plus-2025-12-01, qwen-plus-2025-09-11, qwen-plus-2025-07-28, qwen-plus-2025-07-14, qwen-plus-2025-04-28, qwen-plus-2025-01-25, qwen-plus-2025-01-12, qwen-plus-2024-12-20 — Qwen Plus family
• qwen3.6-flash, qwen3.6-flash-2026-04-16, qwen3.5-flash, qwen3.5-flash-2026-02-23, qwen-flash, qwen-flash-2025-07-28, qwen-flash-character, qwen-turbo, qwen-long, qwen-long-latest, qwen-long-2025-01-25 — Qwen Flash/Turbo/Long family
• qwen-mt-plus, qwen-mt-turbo, qwen-mt-flash, qwen-mt-lite, qwen-plus-character, qwen-plus-character-ja — Translation and character-role models
• qwen3.6-35b-a3b, qwen3.5-397b-a17b, qwen3.5-122b-a10b, qwen3.5-27b, qwen3.5-35b-a3b, qwen3-next-80b-a3b-thinking, qwen3-next-80b-a3b-instruct, qwen3-235b-a22b, qwen3-235b-a22b-thinking-2507, qwen3-235b-a22b-instruct-2507, qwen3-32b, qwen3-30b-a3b, qwen3-30b-a3b-thinking-2507, qwen3-30b-a3b-instruct-2507, qwen3-14b, qwen3-8b — Qwen dense/MoE models
• qwq-plus, qvq-max, qvq-max-2025-08-28, qvq-plus, qvq-plus-2025-08-27, qwen3-coder-plus, qwen3-coder-plus-2025-09-23, qwen3-coder-plus-2025-07-22, qwen3-coder-flash, qwen3-coder-flash-2025-07-28, qwen3-coder-next, qwen3-coder-480b-a35b-instruct, qwen3-coder-30b-a3b-instruct — Reasoning, visual reasoning, and coding specialists
• qwen2.5-omni-7b, qwen3.5-omni-plus, qwen3.5-omni-flash, qwen3-omni-flash, qwen3-omni-flash-2025-10-22, qwen-omni-turbo, qwen3-vl-plus, qwen3-vl-plus-2026-01-25, qwen3-vl-flash, qwen3-vl-flash-2026-01-25, qwen-vl-plus, qwen-vl-max, qwen-vl-ocr, qwen-vl-ocr-latest, qwen-vl-ocr-2025-07-14 — Omni, vision, and OCR chat models

• glm-5.2, glm-5.1, glm-5-turbo, glm-5 — GLM-5 text family
• glm-4.7, glm-4.7-flashx, glm-4.7-flash — GLM-4.7 family
• glm-4.6, glm-4.5, glm-4.5-air, glm-4.5-x, glm-4.5-airx, glm-4.5-flash, glm-4-32b-0414-128k — Previous generation
• glm-5v-turbo, glm-4.6v, autoglm-phone-multilingual, glm-4.6v-flash, glm-4.6v-flashx, glm-4.5v — GLM vision chat family

• kimi-k2.7-code, kimi-k2.7-code-highspeed — Coding-focused Kimi K2.7 models
• kimi-k2.6, kimi-k2.5 — General Kimi K-series models
• moonshot-v1-128k, moonshot-v1-32k, moonshot-v1-8k — Moonshot V1 text models
• moonshot-v1-128k-vision-preview, moonshot-v1-32k-vision-preview, moonshot-v1-8k-vision-preview — Moonshot V1 vision previews

• MiniMax-M3 — Latest M-series language model
• MiniMax-M2.7, MiniMax-M2.7-highspeed — M2.7 family
• MiniMax-M2.5, MiniMax-M2.5-highspeed — M2.5 family
• MiniMax-M2.1, MiniMax-M2.1-highspeed — M2.1 family
• MiniMax-M2 — Earlier agentic reasoning model
• M2-her — Dialogue-focused chat model

• Popular: llama3.1, deepseek-r1, gemma3, qwen3, mistral, phi4, llava
• Any model installed — run ollama list to see available models

• Works with LM Studio, LocalAI, text-generation-webui, vLLM, and any OpenAI-compatible server
• Set base_url to your local server (e.g., http://localhost:1234/v1)

Control how the AI behaves with a three-axis model. Each axis is independent, and presets are named combinations for common workflows.

Modifiers are behavioral overlays that stack on top of any preset:

# CLI flag
openclaudia --mode create
openclaudia --mode safe

# In-session switching
/mode                        # Show current mode and list presets
/mode create                 # Switch to create preset
/mode create +bold           # Create preset with bold modifier
/mode debug +context-pacing  # Debug with pacing
/mode safe +bold +readonly   # Stack multiple modifiers

The mode system integrates with Anthropic's prompt caching: behavioral axes and modifiers are part of the stable prompt prefix (cached across turns), while hooks, memory, and environment info are in the dynamic suffix (reprocessed each turn). Mode switches naturally invalidate the prefix cache.

OpenClaudia includes a built-in adversarial code review system. When enabled, a separate AI model (the "adversary") reviews every response for bugs, security vulnerabilities, and logic errors.

vdd:
  enabled: true
  mode: advisory        # Single-pass review, findings injected as context
  adversary:
    provider: google    # Use a different provider than your builder
    model: gemini-3.1-pro-preview
  static_analysis:
    auto_detect: true   # Automatically runs cargo clippy, cargo test, etc.

Two modes:

• Advisory — Single adversary pass after each response. Findings are displayed and injected into context for the next turn.
• Blocking — Full adversarial loop. The builder must revise until the adversary's findings converge to false positives (confabulation threshold).

Findings include CWE classifications, severity levels (CRITICAL/HIGH/MEDIUM/LOW/INFO), and can automatically create Crosslink issues for tracking.

Configure hooks in .openclaudia/config.yaml to run scripts at key moments:

hooks:
  session_start:
    - hooks:
        - type: command
          command: python .openclaudia/hooks/session-start.py
          timeout: 30

  user_prompt_submit:
    - hooks:
        - type: command
          command: python .openclaudia/hooks/prompt-guard.py

  pre_tool_use:
    - matcher: "Write|Edit"
      hooks:
        - type: command
          command: python .openclaudia/hooks/validate-write.py

• session_start — When a session begins
• session_end — When a session ends
• user_prompt_submit — Before processing user input
• pre_tool_use — Before executing a tool (with matcher for specific tools)
• post_tool_use — After executing a tool
• stop — For iteration/loop mode control

OpenClaudia automatically learns from your coding sessions without any flags or model intervention. A SQLite database (.openclaudia/memory.db) captures knowledge from tool execution signals:

• Coding Patterns — Conventions, pitfalls, and architecture observed from lint output and edit failures
• Error Resolutions — Errors encountered and how they were fixed, matched automatically when subsequent commands succeed
• File Relationships — Files frequently edited together (co-edit tracking), surfaced when you touch related code
• User Preferences — Style and workflow preferences detected from corrections ("no, use tabs") and explicit statements ("always use snake_case")
• Session Continuity — Recent session summaries and activity logs for context across restarts

Knowledge is injected into the model's context automatically — file-specific patterns when you read/edit a file, and preferences in every system prompt. Use /memory commands to inspect what's been learned.

.openclaudia/
├── config.yaml        # Main configuration
├── session/           # Persisted chat sessions
├── memory.db          # Auto-learning memory database
├── hooks/             # Custom hook scripts
├── rules/             # Language-specific rules (*.md)
├── plugins/           # Plugin manifests
├── logs/              # Audit logs
└── vdd/               # VDD session logs (if tracking enabled)

# Development build (includes browser feature by default)
cargo build

# Release build
cargo build --release

# Without browser feature (smaller binary; web_search is unavailable)
cargo build --release --no-default-features

# Run all tests
cargo test

# Run integration tests (tests real tool execution)
cargo test --test integration_tests

# Lint
cargo clippy -- -D warnings

# Run with verbose logging
RUST_LOG=debug cargo run

OpenClaudia is built with:

• tokio — Async runtime
• axum — HTTP server (for proxy mode)
• reqwest — HTTP client
• rusqlite — SQLite for memory
• ratatui — Terminal UI
• rustyline — Line editing with history
• crossterm — Terminal manipulation
• serde — Serialization
• clap — CLI argument parsing
• tracing — Structured logging

Default features (can be disabled with --no-default-features):

• headless_chrome — Headless browser fallback for web_fetch and no-key DuckDuckGo/Bing web search
• scraper — HTML parsing for search result extraction

MIT License — See LICENSE

Built with Rust. Powered by curiosity.