Open-source Claude Code alternative — multi-provider AI coding agent CLI. Single file, 13 providers, zero native deps.

# npm (recommended) npm install -g cloclo # Or run directly npx cloclo # Or clone and run git clone https://github.com/SeifBenayed/claude-code-sdk.git cd claude-code-sdk && node claude-native.mjs

# Login (opens browser, saves token to macOS keychain) cloclo --login # Anthropic (Pro/Max subscription) cloclo --openai-login # OpenAI (ChatGPT Plus/Pro subscription) # Interactive REPL cloclo # Default: Claude Sonnet cloclo -m codex # OpenAI Codex cloclo -m ollama/llama3.2 # Local Ollama # One-shot cloclo -p "explain this code" cloclo -m gpt-5.4 -p "fix the bug in main.js" cat error.log | cloclo -p "explain this error" # Programmatic (NDJSON bridge) echo '{"type":"message","content":"hello"}' | cloclo --ndjson

Switch providers live in the REPL: /model codex, /model sonnet, /model ollama/llama3.2

• Multi-provider: 13 backends, one CLI. Switch mid-conversation.
• Full agent loop: streaming, tool calling, multi-turn, sub-agents
• Built-in tools: Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch, Agent
• Extended thinking: --thinking 8192 for Claude Sonnet/Opus
• MCP integration: --mcp-config servers.json for external tool servers
• Permissions: 6 modes from auto to bypassPermissions
• Session management: resume, checkpoints, rewind
• Memory system: persistent cross-session memory
• Skills & hooks: extensible slash commands, lifecycle hooks (command, webhook, prompt, agent)
• NDJSON bridge: programmatic use from any language
• Ink UI: rich terminal interface with slash menu, status bar, and live output

/model [name] Switch model (live backend switching) /thinking [budget] Toggle extended thinking /compact Compress conversation to save context /permissions <mode> Change permission mode /memory Show saved memories /checkpoints List file checkpoints /rewind Restore files to a checkpoint /sessions List recent sessions /cost Show session cost /webhook Manage webhook hooks /clear New session /login Login to Anthropic /openai-login Login to OpenAI /exit Quit 

-p, --print <prompt> One-shot mode -m, --model <name> Model name or alias --provider <name> Force a specific provider --ndjson NDJSON bridge mode (stdin/stdout) --output <format> Output format: text (default) or json --json Shorthand for --output json --output-version <v> Lock JSON output schema version (default: 1) -y, --yes Skip all permission prompts --timeout <seconds> Global timeout (exit code 5 if exceeded) --login Anthropic OAuth login --openai-login OpenAI OAuth login --api-key <key> Anthropic API key --openai-api-key <key> OpenAI API key --permission-mode <mode> auto|default|plan|acceptEdits|bypassPermissions|dontAsk --max-turns <n> Max agent turns (default: 25) --max-tokens <n> Max output tokens (default: 16384) --thinking <budget> Extended thinking token budget --mcp-config <path> MCP servers config file --resume Resume most recent session --session-id <id> Resume specific session --verbose Debug logging 

Use --json (or --output json) in one-shot mode for machine-readable output:

cloclo -p "what is 2+2" --json

{ "version": "1", "message": "2 + 2 = 4", "model": "claude-sonnet-4-6", "provider": "Anthropic", "usage": { "input_tokens": 42, "output_tokens": 12, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0 }, "stop_reason": "end_turn", "turns": 1, "session_id": "abc123..." }

Pin the schema version with --output-version to protect against future changes.

cloclo --login # OAuth — uses your Pro/Max subscription # or ANTHROPIC_API_KEY=sk-ant-... cloclo # API key — pay-per-token

cloclo --openai-login # OAuth — uses your ChatGPT Plus/Pro subscription # or OPENAI_API_KEY=sk-... cloclo -m codex # API key

# Ollama (install: https://ollama.com) ollama pull llama3.2 cloclo -m ollama/llama3.2 # LM Studio cloclo -m lmstudio/qwen2.5-coder # Custom URL OLLAMA_API_URL=http://gpu-box:11434 cloclo -m ollama/llama3.2

For programmatic use from any language:

→ {"type":"message","content":"search for X","tools":[...]} ← {"type":"ready","version":"...","session_id":"..."} ← {"type":"tool_use","id":"...","name":"Bash","input":{"command":"..."}} → {"type":"tool_result","id":"...","content":"...","is_error":false} ← {"type":"response","content":"Found 3 results...","iterations":2} 

Pass custom tools in the message payload — the SDK calls them via NDJSON and returns results.

cloclo --mcp-config mcp-servers.json

{ "mcpServers": { "my-tools": { "command": "node", "args": ["my-mcp-server.js"] } } }

Connect external CLIs and HTTP services as first-class tools.

cloclo tool install ./my-tool/ # directory with TOOL.json cloclo tool list # shows name, type, state cloclo tool info my-tool # type-specific details cloclo tool test my-tool # binary check + healthcheck cloclo tool remove my-tool

For real CLI products (gh, vercel, ramp, etc.):

{ "name": "github-pr-list", "type": "cli", "description": "List GitHub pull requests", "binary": "gh", "args_template": ["pr", "list", "--json", "number,title,state"], "input_schema": { "type": "object", "properties": {} }, "timeout": 10000, "read_only": true, "parse_mode": "json", "env": ["GITHUB_TOKEN"], "healthcheck": ["gh", "--version"] }

Fields: binary (required), args_template, parse_mode (json/text/lines), read_only (required), env (required env vars), healthcheck (verify command), stdin_template, exit_code_map, success_exit_codes, cwd, timeout.

tool test checks: binary exists → env vars present → healthcheck passes.

For REST APIs and enterprise services:

{ "name": "hedi-fraud-check", "type": "http", "description": "Check if a document is suspicious", "method": "POST", "url": "https://hedi.internal/api/fraud/check", "headers": { "Authorization": "Bearer ${HEDI_TOKEN}" }, "timeout": 10000, "read_only": true, "healthcheck_url": "https://hedi.internal/health", "error_map": { "401": "Auth failed — set HEDI_TOKEN", "503": "Service down" }, "input_schema": { "type": "object", "properties": { "document_text": { "type": "string" } }, "required": ["document_text"] } }

Enterprise features:

• ${ENV_VAR} interpolation in headers and URL — missing vars fail loudly
• error_map — map HTTP status codes to actionable messages
• healthcheck_url — reachability check via tool test
• auth_env — declare which env var holds the auth token

tool test checks: env vars → healthcheck_url reachability (or "No healthcheck configured").

# Unit tests (no API keys needed) npm test # E2E tests (needs API keys) node test-suite.mjs --e2e # Ink UI smoke tests node test-ink-smoke.mjs # Verbose output on failures node test-suite.mjs --verbose

Single-file SDK (claude-native.mjs, ~7500 lines) with provider-pluggable core. AgentLoop reads capabilities from the provider contract — never checks provider names. All provider-specific logic lives in the provider definition.

AgentLoop (streaming, tools, permissions, hooks) ├── AnthropicClient → /v1/messages ├── OpenAIClient → /v1/chat/completions ├── OpenAIResponsesClient → /v1/responses └── OpenAI-compat → Gemini, DeepSeek, Mistral, Groq, local 

Every provider implements: detect(), createClient(), resolveAuth(), capabilities.

MIT