Key takeaways
- Single-file architecture (one index.js) makes deployment and customization trivially simple
- Built on official Claude Agent SDK — first-party integration, not reverse-engineered
- Smart automation features (personality adaptation, mood tracking, skill discovery) differentiate from minimal alternatives
FAQ
What is BabyClaw?
A lightweight, single-file alternative to OpenClaw built on the official Claude Agent SDK. Runs on a VPS and is controlled entirely through Telegram. As of June 2026 the project is dormant, with no commits since February 17, 2026.
How much does BabyClaw cost?
Free and open source (MIT license). You pay for your VPS (~$5-20/mo) and Claude subscription or API costs.
Who competes with BabyClaw?
OpenClaw (full-featured), Pi (minimal core), NullClaw (smallest footprint), ZeptoClaw (balanced security).
Executive Summary
Status (June 2026): Stalled. The repository is not archived, but the last commit landed February 17, 2026 — eight days after the project was created, and four days before this profile was first written. Nearly four months have passed with no further activity. Stars grew from ~1 to 15 over that period, but with no commits, no releases, and no community discussion found anywhere, this looks like a published weekend project rather than a maintained one.[1]
BabyClaw is a lightweight, single-file alternative to OpenClaw built on the official Claude Agent SDK. It runs on a VPS and is controlled entirely through Telegram — send text, voice messages, or images, and Claude responds. Unlike minimal alternatives focused purely on size, BabyClaw includes smart automation features like personality adaptation, mood tracking, and skill discovery.
Funding: not publicly disclosed (solo side project; no evidence any was raised).
| Attribute | Value |
|---|---|
| Developer | Yogesh (@yogesharc) |
| Language | JavaScript (Node.js) |
| License | MIT |
| GitHub Stars | 15 ★ (as of June 2026) |
| Last commit | February 17, 2026 |
| Architecture | Single file (index.js) |
Naming note: An unrelated project also called BabyClaw (babyclaw/babyclaw, TypeScript, at babyclaw.org) appeared on February 21, 2026 with the tagline "same lobster spirit, ~5% of the complexity." It is a different codebase by a different author — and is itself dormant since February 23, 2026. This profile covers yogesharc/babyclaw only.[2][3]
Product Overview
BabyClaw positions itself as "Claude Code on a VPS, controlled from Telegram." The core insight: most personal AI agent users don't need 50+ messaging integrations — they need one reliable channel that works everywhere. Telegram provides mobile + desktop + web access with rich media support.
The single-file architecture is intentional. One index.js means you can read the entire codebase, understand exactly what's happening, and modify it freely. No framework abstractions, no database dependencies, no complex build systems.
Key Capabilities
| Capability | Description |
|---|---|
| Voice messages | Send voice in Telegram, auto-transcribed via Whisper |
| File delivery | Ask Claude for any file and it sends directly to Telegram |
| Persistent memory | Conversations auto-save, context survives session resets |
| Personality adaptation | Mirrors your communication style over time (cron updates every 3 days) |
| Mood tracking | Silently logs mood shifts, tracks patterns |
| Skill discovery | Daily scan suggests relevant skills from ClawdHub/skills.sh |
| Pattern detection | Finds automation opportunities, suggests cron jobs |
| Browser capabilities | Opens URLs, interacts with pages via agent-browser |
Memory System
BabyClaw has a three-layer memory architecture:
- Session continuity — Full context within a session, persists across bot restarts
- History — Long-term memory saved as dated markdown files when starting new sessions
- Recent conversations — Rolling list of last 50 history entries loaded into every session
Additional tracking includes mood logs (detecting emotional patterns) and open threads (unresolved topics that resurface when relevant).
Technical Architecture
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Telegram │────▶│ VPS │────▶│ Claude SDK │
│ (mobile) │◀────│ index.js │◀────│ (API) │
└─────────────┘ └─────────────┘ └─────────────┘
│
┌──────┴──────┐
│ workspace │
│ + history │
│ + crons │
└─────────────┘
| Aspect | Detail |
|---|---|
| Deployment | VPS (any provider, Ubuntu recommended) |
| Runtime | Node.js + tmux |
| Model | Claude (Sonnet/Opus, switchable on the fly) |
| Voice transcription | OpenAI Whisper API |
| Browser | agent-browser for web interaction |
Built-in Crons
- Pattern detector — Analyzes conversations, suggests automation opportunities
- Skills suggester — Searches ClawdHub/skills.sh for relevant skills
- Personality updater — Updates Claude's communication style every 3 days
- Daily digest — End-of-day summary sent to Telegram
Strengths
- Radical simplicity — One file, no framework, no database. You can read the entire codebase in an hour
- Official SDK — Built on Claude Agent SDK, not reverse-engineered APIs
- Smart automation — Personality adaptation and mood tracking go beyond basic assistants
- Voice-first option — Send voice messages naturally, transcription is automatic
- Self-improving — Crons continuously analyze usage and suggest improvements
Cautions
- Stalled development — No commits since February 17, 2026; the entire commit history spans eight days
- Telegram-only — No WhatsApp, Discord, Slack, or iMessage support
- Very early and tiny — 15 GitHub stars, 1 fork, single developer
- VPS required — More setup than managed alternatives
- Limited integrations — No MCP support, no skills ecosystem yet
- No isolation — Claude has full access to workspace and VPS
- Name collision — An unrelated, also-dormant "BabyClaw" project (babyclaw.org) makes discovery and search confusing[2]
Pricing & Licensing
| Component | Cost |
|---|---|
| BabyClaw | Free (MIT license) |
| VPS | ~$5-20/mo (any provider) |
| Claude | Subscription (Pro/Max) or API usage |
| Whisper | OpenAI API usage (optional, for voice) |
Total: ~$25-50/mo depending on usage and VPS choice.
Competitive Positioning
| Alternative | BabyClaw Advantage |
|---|---|
| OpenClaw | Simpler (1 file vs 430K+ LOC), easier to understand and modify |
| Pi | More features (voice, crons, mood tracking) vs Pi's minimal 4-tool core |
| NullClaw | Smarter (auto-improving crons) vs pure efficiency focus |
| ZeptoClaw | Simpler setup, no Rust compilation required |
When to Choose BabyClaw
- You want a single-file codebase you can fully understand
- Telegram is your primary messaging app
- You value smart automation (personality adaptation, mood tracking)
- You're comfortable with VPS deployment
When to Choose Alternatives
- OpenClaw — Need 50+ messaging integrations
- Pi — Want minimal core to build on
- NullClaw/ZeroClaw — Security or efficiency is top priority
- Lindy — Want managed solution with zero setup
What Developers Say
As of June 11, 2026, there is no community discussion of BabyClaw to report: searches of Hacker News, Reddit, and X surfaced no threads, reviews, or substantive mentions of the project. For a 15-star repository with one fork and an eight-day commit history, that silence is expected — the only public footprint is the repository itself.[1]
Bottom Line
BabyClaw remains a clean illustration of an idea — Claude Code on a VPS, controlled from Telegram, in one readable file — but as of June 2026 it is a snapshot, not a project. Development stopped on February 17, 2026, eight days after it began, and nothing has shipped since. The code still works as described and the single-file architecture is still genuinely instructive, but anyone adopting it should treat it as unmaintained starter code to fork, not software with a future.
Recommended for: Developers who want a small, readable reference implementation of a Telegram-controlled Claude agent to fork and own themselves.
Not recommended for: Anyone expecting maintenance, updates, multiple messaging channels, or managed hosting.
Outlook: Effectively dormant. The single-file philosophy did spread — an unrelated namesake project appeared within two weeks promising "~5% of the complexity" of OpenClaw (and promptly went dormant too) — but this particular implementation shows no signs of further development. If activity hasn't resumed by the next refresh, this profile is a candidate for retirement.[1][2]
Research by Ry Walker Research • methodology