Any Model, One Agent — LLM Harness that unifies Claude CLI, Codex CLI, and Gemini CLI under a single proxy with a web dashboard.

AniClew sits between your coding CLI tools and LLM providers, giving you multi-provider routing, per-project management, multi-agent orchestration, and a visual control plane.

Chat (Thinking model)	Project Browser

KAIROS Daemon	Observability

• 7 providers: Anthropic, OpenAI, Gemini, Groq, Ollama, GitHub Copilot, z.ai (Grok)
• Auth passthrough: CLI tools send their own API keys through AniClew transparently
• Runtime switching: Change provider/model without restarting
• Smart retry: Exponential backoff with jitter, 529 fallback model switching
• Smart router: Auto-route requests by role (coding, review, chat)

• Tool-using agent: Bash, Read, Write, Edit, Glob, Grep
• Thinking model support: qwen3, DeepSeek-R1 reasoning in collapsible blocks
• 23 security validators: Shell injection detection, dangerous path blocking, sed/jq execution prevention
• 60+ read-only allowlist: Per-command flag validation for safe auto-approval
• Parallel tool execution: Read-only tools run concurrently, write tools serial
• Context auto-compaction: LLM-based summarization with snip fallback + circuit breaker

• Wave execution: Topological sort (Kahn's algorithm) for dependency-based parallelism
• File ownership: Hard enforcement at ExecuteTool level
• 6 agent types: Explorer, Researcher, Planner, Coder, Reviewer, Tester
• Mailbox: File-based inter-agent messaging with locking + broadcast
• Message router: Live channel delivery with mailbox fallback
• Plan mode: Explore -> Design -> Approve -> Implement lifecycle
• Worktree isolation: Git worktree per agent for conflict-free parallel work
• Idle detection: Callback chain with timestamp tracking

• Multi-project: Register, switch, remove projects via folder browser
• File tree: Recursive tree with click-to-view file content
• Session isolation: Per-workspace chat history filtering
• Auto-detection: Go, Node, Python, Rust, Java, .NET framework detection

• Background agent: 2-minute tick cycle with cron scheduling
• Cron parser: 5-field expressions + presets (@hourly, @daily, @every 5m)
• Git Watch: Auto-monitors git status, detects changes, logs summaries
• Notifications: SSE real-time stream + webhook integration
• Per-project: Tasks and memory isolated per workspace

• Request tracing: Per-request provider, model, latency, tokens, cost (JSONL persistence)
• Metrics: Average/P95 latency, error rate, requests/min, per-provider breakdown
• Stream watchdog: 90s idle timeout with context abort
• Response quality: Thumbs up/down feedback with per-model scoring
• Prompt cache: Strategy tracking, break detection, savings estimation

• Hook system: CLI-aware loading (Claude/Codex/Gemini settings), pre/post tool use
• 6-level permission cascade: CLI flags > policy > local > user > project > defaults
• Immutable snapshots: Captured at session start, no mid-session drift
• Denial tracking: Auto-persist deny rule after 3 consecutive denials
• Permission persistence: Allow/deny rules saved to .claude/settings.json

• Token auth: Optional accessToken in config
• Path sandboxing: Tools restricted to workspace directory
• BashTool: Quote-aware parser, compound command splitting, auto-backgrounding
• Exit code semantics: grep 1 = no match (not error), diff 1 = files differ

• Plugin system: JSON manifest with tools, hooks, commands, agent types
• MCP integration: Stdio client with timeout, health check, auto-reconnect
• Bridge mode: HTTP remote control for IDE/script integration
• Session memory: Disk-backed large result storage (saves tokens)

curl -fsSL https://raw.githubusercontent.com/Dannykkh/Ani-Clew/main/install.sh | bash

# With Ollama (local LLM)
docker compose up -d

# AniClew only (bring your own LLM)
docker run -p 4000:4000 -v $(pwd):/workspace ghcr.io/dannykkh/aniclew

git clone https://github.com/Dannykkh/Ani-Clew.git && cd Ani-Clew
make all    # builds frontend + backend
./aniclew   # interactive provider select

Go to Releases and download for your platform.

# Start with Ollama (local, free)
./aniclew -provider ollama -model qwen3:14b

# Start with OpenAI
OPENAI_API_KEY=sk-... ./aniclew -provider openai -model gpt-4o

# Start with Anthropic
ANTHROPIC_API_KEY=sk-ant-... ./aniclew -provider anthropic -model claude-sonnet-4-6-20250217

Browser opens at http://localhost:4000/app.

# Claude CLI → AniClew → any model
ANTHROPIC_BASE_URL=http://localhost:4000 claude

# Codex CLI → AniClew → any model
OPENAI_BASE_URL=http://localhost:4000 codex

~/.claude-proxy/config.json:

{
  "port": 4000,
  "defaultProvider": "ollama",
  "defaultModel": "qwen3:14b",
  "accessToken": "",
  "projects": [
    { "path": "/path/to/project", "name": "My Project" }
  ],
  "providers": {
    "ollama-remote": { "baseUrl": "http://192.168.1.100:11434" }
  }
}

CLI Tools (Claude/Codex/Gemini)
        |
        v
  +-----------+     +---------+
  |  AniClew  | <-> | Web UI  |
  +-----------+     +---------+
   |    |    |
   v    v    v
Anthropic OpenAI Ollama ...
   |
   +-- Agent Loop (tools, hooks, permissions)
   +-- KAIROS Daemon (cron, git-watch)
   +-- Team (waves, mailbox, worktree)
   +-- Observability (traces, metrics, feedback)
   +-- Plugins (tools, hooks, commands)

• 17,871 lines Go backend
• 214 tests across 4 packages
• 95% technical fidelity
• 11 runtime-verified features

MIT