Key takeaways
- Renamed to safestclaw and pivoted to LLM-optional — core features stay rule-based and zero-cost, but Ollama, OpenAI, Anthropic, Google, Mistral, and Groq can now be plugged in for research, blogging, and code tasks
- Zero API cost by default — traditional ML (VADER, spaCy, sumy, Whisper, Piper) handles the core; prompt-injection-immune in no-LLM mode
- Grew from 69 to 277 GitHub stars between February and June 2026, with active releases (v4.2.4, May 2026) and a pip-installable package
FAQ
What is safeclaw?
An OpenClaw alternative, renamed safestclaw in 2026, that uses traditional ML pipelines by default instead of LLMs. Zero API cost by default, no prompt injection in no-LLM mode, deterministic output. Optional LLM providers can be enabled for research and coding.
How much does safeclaw cost?
$0 by default — no LLM API costs unless you opt into a cloud provider. Optional ML features add ~50MB-2GB disk. MIT licensed.
Who competes with safeclaw?
OpenClaw (LLM-based), NullClaw (minimal LLM), Pi (lightweight LLM).
Executive Summary
safeclaw — renamed safestclaw since this profile was first published — started with a radical premise: what if your AI assistant didn't need an LLM at all? Using VADER sentiment analysis, spaCy NER, sumy summarization, YOLO object detection, Whisper STT, and Piper TTS, it delivers most of OpenClaw's day-to-day functionality with zero API costs and no prompt injection risk in its default mode [safeclaw-github].
The project has since softened its purity stance: as of mid-2026, LLM integration is optional — Ollama, OpenAI, Anthropic, Google, Mistral, and Groq can be plugged in for blogging, research synthesis, and code tasks, with an LLM-based NLU bridge that is disabled by default [safestclaw-readme]. The pitch is now "zero cost by default, AI when you want it."
| Attribute | Value |
|---|---|
| Language | Python |
| License | MIT |
| GitHub Stars | 277 ★ (as of June 2026, up from 69 in February) |
| LLM Required | No (optional since v4.x) |
| Monthly Cost | $0 by default |
| Install | pip install safestclaw (v4.2.x) [safestclaw-pypi] |
Status (June 2026): Active. Latest push May 18, 2026; release v4.2.4 shipped May 9, 2026; 34 forks [safeclaw-github] [safestclaw-releases].
Key Capabilities
| Capability | Description |
|---|---|
| Voice | Whisper STT + Piper TTS, completely offline |
| Summarization | LexRank, TextRank, LSA, Luhn algorithms |
| Sentiment | VADER lexicon-based analysis |
| RSS/News | 50+ preset feeds, 8 categories |
| Social media | Twitter, Mastodon, Bluesky summaries |
| IMAP/SMTP integration | |
| Calendar | ICS files, CalDAV |
| Documents | PDF, DOCX, HTML, Markdown |
| Web crawling | Async with depth limits |
| Smart home | Device control, Bluetooth, network scanning |
| Vision | YOLO object detection + OCR (optional) |
| Blog publishing | WordPress, Joomla, SFTP (new since Feb 2026) |
| Research pipeline | arXiv, Semantic Scholar, Wolfram Alpha; LLM-optional synthesis (new) |
| Security scanning | bandit, semgrep, trivy integrations (new) |
| MCP | Model Context Protocol support (new) |
| Channels | CLI, Telegram, webhooks, local web UI; Discord/Slack announced |
[safestclaw-readme]
What safestclaw Can't Do (without an LLM)
| Feature | safestclaw (no-LLM mode) | OpenClaw |
|---|---|---|
| Free-form chat | ❌ (rule-based intent parsing) | ✅ |
| Creative writing | ❌ (optional LLM for blogging) | ✅ |
| Autonomous multi-step tasks | ❌ | ✅ |
| Self-writing skills | ❌ | ✅ |
| Browser automation | ❌ | ✅ |
Cost Comparison
| Metric | safestclaw (default) | OpenClaw |
|---|---|---|
| Monthly cost | $0 | $100-$3,600+ |
| Prompt injection risk | None (no-LLM mode) | Yes |
| Works offline | Yes | No |
| Deterministic output | Yes | No |
Strengths
- Zero cost by default — No LLM API bills unless you opt in
- No prompt injection in no-LLM mode — Can't inject what doesn't exist
- Deterministic core — Same input always produces same output
- Privacy — Local by default, external only when you ask
- Offline — Core features work without internet
- Voice — Full STT/TTS completely offline
- Active development — 4x star growth and steady releases February–May 2026 [safeclaw-github]
Cautions
- Identity drift — The rename (safeclaw → safestclaw) and the addition of optional LLM providers blur the original "no LLM, period" pitch; injection immunity only holds in the default mode [safestclaw-readme]
- Limited reasoning — Rule-based core, not intelligent
- Small community — 277 stars and a single primary maintainer; the HN launch drew effectively no third-party engagement [safeclaw-hn-launch]
- Feature gaps — No browser automation, no autonomous multi-step tasks
- Structured input — Core mode requires command-style phrasing
What Developers Say
No substantive third-party commentary found as of June 2026. The project's February 2026 Hacker News launch received 2 points and a single comment — from the author's own account, which summarized the pitch: "No AI, no apis, it all actually stays on your machine" [safeclaw-hn-launch]. Star growth (69 → 277) suggests quiet interest, but there are no independent reviews, benchmarks, or notable user testimonials to cite yet.
Bottom Line
safestclaw remains a fascinating experiment — how much assistant can you build without an LLM? — but it has quietly become a hybrid: a deterministic, zero-cost core with opt-in LLM providers for research, blogging, and code. That pragmatic pivot makes it more useful and more comparable to its LLM-based category peers, at the cost of its purist security pitch. It's still not a replacement for OpenClaw-style autonomous agents, but for cost-conscious users who want voice, news, email, documents, and basic automation with a $0 default bill, it delivers — and it's the rare project in this category that got more alive since February.
Recommended for: Users who want basic assistant features with zero default API costs, deterministic behavior, and the option to bolt on an LLM later.
Not recommended for: Anyone who needs natural conversation by default, creative tasks, or autonomous reasoning.
Research by Ry Walker Research • methodology