shibaclaw 0.0.25

A lightweight personal AI assistant framework

ShibaClaw 🐕

Security-first AI agent with built-in WebUI, native provider support, and hardened tools.

---

ShibaClaw is a security-first AI agent that runs in your terminal or in a browser-based WebUI. Instead of assuming the surrounding app will handle safety, it builds it into the core: install-time CVE auditing, randomized tool-output wrapping against prompt injection, SSRF and DNS rebinding protection, shell hardening, workspace sandboxing, and token auth. You still get the practical pieces you need for daily use: WebUI & onboarding, 22 LLM providers, built-in file tools, long-term memory, 11 chat channels, cron, heartbeat, skills, and MCP support.

Security, Built In

These are the defenses that are often left to app glue code or external proxies. In ShibaClaw they are part of the framework itself.

Layer	Built in by default	Why it matters
Install-time audit	Audits pip and npm installs before execution; blocks critical/high CVEs	Catches risky dependencies before they land in the environment
Prompt-injection wrapping	Wraps every tool result in a randomized <tool_output_...> boundary and sanitizes closing tags	Untrusted pages and files stay data, not instructions
Shell hardening	20+ deny patterns, escape normalization (\x.., \u....), internal URL detection	Blocks common destructive or obfuscated commands
Network guard	SSRF filtering, redirect revalidation, DNS-rebinding-safe resolution	Prevents web tools from pivoting into localhost or private networks
Workspace sandbox	File tools and the WebUI file browser stay inside the configured workspace	Reduces traversal and accidental host-wide access
Access control	Bearer token auth, constant-time token checks, channel allowlists, optional sender rate limiting	Safer when the agent is exposed beyond a local shell

pip install shibaclaw
shibaclaw web               # open http://localhost:3000 → onboard wizard starts

---

Quick Start

Docker

git clone https://github.com/RikyZ90/ShibaClaw.git && cd ShibaClaw
docker compose up -d --build
docker exec -it shibaclaw-gateway shibaclaw print-token

Open http://localhost:3000 — paste the token if auth is enabled, then complete the onboard wizard in the browser.

pip

pip install shibaclaw
shibaclaw web --port 3000

Open http://localhost:3000 and complete the onboard wizard. Prefer the terminal? shibaclaw onboard runs the same guided setup from the CLI.

---

WebUI

  

The WebUI is built-in — no separate frontend or Node.js required.

• Chat — multi-session conversations with live streaming of tool calls, thinking blocks, and elapsed time
• Agent Profiles — switch personas per session (Hacker, Builder, Planner, Reviewer) with dynamic avatars
• File browser — browse, view, and edit workspace files in-browser (sandboxed to workspace)
• Voice — speech-to-text via OpenAI-compatible audio APIs and browser-native TTS
• Settings — configure agent, provider, tools, MCP servers, channels, skills, and OAuth from a single panel
• Onboard wizard — guided first-time setup: pick a provider, enter API key or start OAuth, choose a model
• Context viewer — inspect the full system prompt and token usage breakdown
• Gateway monitor — health check and one-click restart
• Auto-update — checks GitHub releases every 12h, notifies in the UI and on all active channels
• Responsive — works on desktop and mobile

Agent Profiles

Switch the agent's personality on-the-fly without losing context. Each profile overrides the system prompt (SOUL.md) while keeping model, memory, and tools shared. Profiles are per-session — run a security audit in one tab and plan architecture in another.

Built-in profiles: Default · Builder · Planner · Reviewer · Hacker (elite security expert with 50+ tool recommendations, OWASP/MITRE/NIST methodologies, CVSS scoring, and a custom cyber-shiba avatar).

Create your own profiles interactively — the agent walks you through defining the persona and saves everything automatically.

---

Features

Memory & Workflow

• Three-level memory — USER.md (personal profile), MEMORY.md (operational facts), HISTORY.md (timestamped session archive with TF-IDF + recency search)
• Proactive learning — every N messages the agent silently consolidates new learnings into memory, without interrupting the conversation
• Focused background delegation — the spawn tool can offload a specific task and report back into the same session when done
• Advanced reasoning — supports extended thinking (Anthropic), reasoning effort (OpenAI o-series), and DeepSeek-R1 chains

Tools

Tool	What it does
exec	Shell commands with 20+ deny-pattern guards, encoding normalization, and CVE scanning
read_file / write_file / edit_file	Paginated reads, fuzzy find-and-replace, auto-created parent dirs
web_search	Brave, Tavily, SearXNG, Jina, or DuckDuckGo (fallback, no key needed)
web_fetch	HTTP fetch with SSRF protection, DNS rebinding defense, and redirect validation
memory_search	Ranked search over session history (TF-IDF + recency + importance scoring)
message	Cross-channel messaging with media attachments
cron	Schedule one-time or recurring jobs (cron expressions, intervals, ISO dates, timezone-aware)
spawn	Optional background worker for a focused task; reports back to the main session when done
MCP	Connect any MCP server (stdio, SSE, or streamable HTTP) — tools auto-registered as mcp_<server>_<tool>

Channels

Telegram · Discord · Slack · WhatsApp · Matrix · Email · DingTalk · Feishu · QQ · WeCom · MoChat

All channels route through the same message bus. WhatsApp uses a Node.js bridge (Baileys) for QR-based linking.

Skills

8 built-in skills (GitHub, weather, summarize, tmux, cron reference, memory guide, skill-creator, ClawHub browser). Skills are Markdown files with YAML frontmatter and optional scripts — create your own or install from ClawHub. Pin frequently-used skills to load them on every conversation.

Automation

• Cron service — persistent, timezone-aware scheduled jobs stored in jobs.json. Supports every, cron, and at schedules. Overdue jobs fire on startup.
• Heartbeat — periodic wake-up reads HEARTBEAT.md, asks the LLM if there's work to do (via structured tool call, not free-text parsing), and executes if needed.

---

Security Policy

The table above is the operational summary. The full disclosure process, supported versions, and defense-in-depth notes live in SECURITY.md.

---

Supported Providers

ShibaClaw uses native SDKs (no LiteLLM proxy) and auto-detects the right provider from the model name.

API Key

Provider	Env Variable
OpenAI	OPENAI_API_KEY
Anthropic	ANTHROPIC_API_KEY
DeepSeek	DEEPSEEK_API_KEY
Google Gemini	GEMINI_API_KEY
Groq	GROQ_API_KEY
Moonshot	MOONSHOT_API_KEY
MiniMax	MINIMAX_API_KEY
Zhipu AI	ZAI_API_KEY
DashScope	DASHSCOPE_API_KEY

Gateway / Proxy

OpenRouter · AiHubMix · SiliconFlow · VolcEngine · BytePlus — auto-detected by key prefix or api_base.

Local

Ollama (http://localhost:11434) · vLLM · any OpenAI-compatible endpoint.

OAuth

Provider	Flow	Setup
GitHub Copilot	Device flow, auto token refresh	shibaclaw provider login github-copilot or WebUI Settings
OpenAI Codex	PKCE browser flow	shibaclaw provider login openai-codex or WebUI Settings

---

Architecture

Docker Compose

Service	Role	Default Port
shibaclaw-gateway	Core agent loop, message bus, channel integrations	19999
shibaclaw-web	WebUI (Starlette + Socket.IO), cron service	3000

Both share the ~/.shibaclaw/ volume (config, workspace, memory, cron jobs, media cache).

Single-process mode

shibaclaw web runs agent + WebUI + cron in a single process — no gateway container needed.

Stack

Layer	Technology
Server	Uvicorn → Starlette (ASGI) + python-socketio
Real-time	Socket.IO (WebSocket primary, polling fallback)
Frontend	Vanilla JS · Marked.js · Highlight.js
Sessions	JSONL append-only per session (cache-friendly for LLM prompt prefixes)

Resource usage

Component	RAM
Gateway	< 200 MB
WebUI	< 200 MB
Total (Docker)	< 400 MB

---

CLI Reference

shibaclaw web               # Start WebUI (agent + cron in-process)
shibaclaw gateway            # Start gateway only (for Docker split)
shibaclaw onboard            # CLI-based first-time setup wizard
shibaclaw agent -m "Hello"   # One-shot message via terminal
shibaclaw agent              # Interactive REPL with history
shibaclaw status             # Provider, workspace, OAuth health check
shibaclaw print-token        # Show WebUI auth token
shibaclaw channels status    # List enabled channels
shibaclaw provider login <p> # OAuth login (github-copilot, openai-codex)

---

Latest — v0.0.25

• Agent Profiles — per-session personas with dynamic avatars (Default, Builder, Planner, Reviewer, Hacker)
• Hacker profile — elite security expert with 50+ tools, OWASP/MITRE methodologies, CVSS scoring
• Custom profile creation — agent-guided interactive persona builder
• Profile startup sync — templates auto-synced, corrupted manifests auto-repaired

→ Full history in CHANGELOG.md

---

Troubleshooting

Problem	Try
General status check	shibaclaw status
Container logs	docker logs shibaclaw-gateway / docker logs shibaclaw-web
WebUI won't connect	Check token with shibaclaw print-token, verify port binding
Provider errors	shibaclaw status shows API key and OAuth state
Security policy	SECURITY.md

---

Contributing

See CONTRIBUTING.md — PRs welcome.

Channels are extensible via Python entry points (shibaclaw.integrations). Skill creation is documented in docs/CHANNEL_PLUGIN_GUIDE.md and the built-in skill-creator skill.

---

Credits

Inspired by NanoBot by HKUDS — MIT License.

---

If you like ShibaClaw and want to help it grow:
⭐ Drop a star — 🐛 Open an issue — 🔧 Send a PR
contributions of any size are welcome