PhoneClaw turns phones into local AI agent runtimes with on-device models, native mobile Skills, LiveLand, and optional Mac Gateway inference.
1.1k
Stars
155
Forks
1
Open issues
2
Contributors
AI Analysis
PhoneClaw transforms iPhones into local AI agent runtimes that execute on-device tasks through native mobile Skills (calendar, reminders, contacts, health data, image understanding) without cloud dependency. It runs Gemma 4 and MiniCPM-V models locally via LiteRT, with optional Mac Gateway inference over local networks. Best suited for iOS developers and privacy-conscious users who need AI agents deeply integrated with device capabilities; not for those requiring cloud-scale models or cross-p...
Inferred from signals mentioned in the README (tests, CI, type safety) — not a review of the actual code.
AI's overall editorial judgment — not an average of the bars above, can weigh other factors too.
On-device iPhone AI agent powered by Gemma 4, handles personal tasks offline without cloud dependency
PhoneClaw is a locally-running AI agent for iOS that performs inference and task execution entirely on-device using Gemma 4 and MiniCPM-V models. It integrates with iPhone system features (calendar, reminders, contacts, health data) through a file-based skill system, supports real-time voice interaction and camera input, and offers optional Mac-based remote inference via local network. Primary audience is privacy-conscious iOS users seeking offline AI capabilities for personal task automation.
Project created April 2026 as a response to demand for on-device iPhone AI agents. Rapidly iterated from v1.0 (basic chat) through v1.4.0 (multimodal support) to current state (agentic skills, LIVE mode, Mac gateway). Appears driven by Gemma model availability and iOS on-device ML infrastructure maturity.
Gained 1,110 stars in ~3 months (April–June 2026), with 43 stars in the most recent week. Growth trajectory suggests early-adopter traction within privacy-focused and iOS developer communities. TestFlight availability (launched ~June 5) and frequent feature releases (LiveLand, Mac gateway, health analytics) have sustained momentum. Comparable peer projects (PalmClaw: 1,104 stars; OpenOmniBot: 1,824 stars) suggest this category is nascent but active.
Adoption not verified through public deployment data, enterprise partnerships, or usage metrics. TestFlight availability suggests beta testing is underway. No documentation of production user counts, retention rates, or organization-level adoption. GitHub stars (1,110) and forks (149) indicate developer interest but do not confirm end-user production usage.
Likely uses a modular skill-based architecture where capabilities (calendar, reminders, health) are defined in Markdown files (SKILL.md) rather than compiled code, enabling non-engineer extensibility. Based on README, employs Swift for iOS frontend, integrates Gemma 4 (via likely MLX or Core ML runtime) and MiniCPM-V for multimodal inference. Includes memory-aware context management, KV cache optimization across conversation turns, and dynamic budget allocation based on available device RAM. Mac gateway component (PhoneClaw Gateway) appears to be a separate Swift/macOS binary using Bonjour for local discovery.
Not documented in README. No mention of unit tests, integration tests, or CI/CD practices.
Last push June 23, 2026 (6 days before evaluation date). High commit frequency evident from detailed changelog spanning April 4–June 23 with updates every few days. Active issue and feature request channels mentioned. Maintenance appears active and responsive; however, project is only 3 months old, so long-term sustainability cannot be assessed.
ADOPT IF: you are an iOS user prioritizing offline AI for personal task automation (calendar, reminders, health tracking), are comfortable with beta-stage software, and can accept limitations of on-device inference (latency, model size constraints). AVOID IF: you require production-grade stability, need cross-platform support, or demand real-world commercial deployment evidence. MONITOR IF: you are evaluating on-device AI agent patterns in mobile ecosystems and want to track whether the file-based skill system and memory optimization techniques mature into industry standards.
Independent dimensions
Mainstream potential
4/10
Technical importance
7/10
Adoption evidence
2/10
- Project age (3 months) means long-term maintenance, security patching, and API stability are unproven. Discontinuation risk exists if creator deprioritizes or iOS platform policies shift.
- On-device inference on iPhone is memory-constrained; README acknowledges E4B model requires CPU-only inference on sideloaded IPA, E2B model recommended as more stable. Performance variability across device generations likely.
- Adoption not verified in production; TestFlight beta stage means real-world error rates, user retention, and support burden unknown.
- Mac gateway feature requires local network pairing, adding operational complexity and limiting utility to home/office environments; security model for local Bonjour discovery not detailed in README.
- File-based skill system promises extensibility but test coverage, validation, and rollback procedures for malformed or adversarial skills not addressed in available documentation.
PhoneClaw likely will stabilize into a niche tool for privacy-conscious iOS power users and developers over the next 6–12 months. If Apple's on-device ML infrastructure matures and Gemma/MLX runtimes improve, it may influence broader iOS agent tooling; otherwise, it will remain a specialized alternative to cloud-based agents. Viability depends on sustained creator engagement and absence of iOS policy restrictions.
Newsletter
Get analyses like this every Monday
Free weekly digest of the most interesting open-source discoveries.
Languages
Information
- Language
- Swift
- License
- NOASSERTION
- Last updated
- 2d ago
- Created
- 3mo ago
- Analyzed with
- anthropic/claude-haiku-4-5
Stars over time
Contributors over time
Top 100 contributors only — repos with more will plateau at 100.
Open issues
Open pull requests
No open pull requests.
Similar repos
ModalityDance/PalmClaw
PalmClaw is a private AI assistant application for Android phones that runs...
nicedreamzapp/claude-code-local
Claude Code Local enables running Claude Code 100% on-device using local AI...
| Repository | Stars | Week Δ | Language | Score | Updated |
|---|---|---|---|---|---|
|
kellyvv/PhoneClaw
this
|
1.1k | +9 | Swift | 8/10 | 2d ago |
|
ModalityDance/PalmClaw
|
1.1k | — | Kotlin | 7/10 | 3w ago |
|
nicedreamzapp/claude-code-local
|
2.9k | — | Python | 8/10 | 3w ago |
 SimonSchubert/Kai
|
1k | — | Kotlin | 7/10 | 3d ago |
 openclaw/openclaw
|
382.4k | — | TypeScript | 8/10 | 3h ago |
 ValueCell-ai/ClawX
|
7.5k | — | TypeScript | 7/10 | 19h ago |
Similar on-device agent for Android with comparable star count (1,104). PhoneClaw targets iOS exclusively; architectural parity unclear from limited metadata.
Multiplatform agent (Dart) with higher star count (1,824). Likely broader platform coverage but less detailed README than PhoneClaw; unclear which offers deeper OS-level integration.
Web-based agent with vastly higher star count (380,916), indicating different positioning. Not directly comparable due to platform and architecture differences.
PhoneClaw can optionally delegate to Ollama via Mac gateway, positioning itself as a iOS UX layer rather than an inference engine replacement.
