OpenOmniBot on-phone Claw — GitHub omnimind-ai avatar | BestClaw

OpenOmniBot Review: On-phone on-device Claw agent

OmniMind AI

An executable agent on your phone — understand/decide/execute/reflect. For early adopters accepting mobile permission risk; BestClaw watchlist positioning.

Review updated: June 14, 2026 · Methodology version aligned with BestClaw rankings

6.7/10

BestClaw composite (28 dimensions)

#25 Unified ranking this cycle

On-phoneTerminal opsClaw agent

Overview

OpenOmniBot (omnimind-ai/OpenOmniBot, Mar 2026) is an on-device Android agent mixing Kotlin and Flutter around an understand → decide → execute → reflect loop — not a chat wrapper. Vision models drive UI automation plus Alpine shell, browser, MCP and system tools (alarms, calendar, audio).

BestClaw keeps it on the watchlist: ~1.5k GitHub stars, still on 0.5.x prerelease, few production war stories. Mobile permission surfaces (accessibility, screen capture, files) are harder than desktop Claws — methodology Ecosystem 5.5 and Features 6.5 mark an early shape, not a final verdict.

Early adopters accept device-loss and misfire risk to automate phones. Unlike PicoClaw minimal servers or LightClaw desktops, OpenOmniBot bets on mobile-native UX.

Before enterprise use, compare with desktop/server OpenClaw on A/B comparison and confirm MDM/compliance tolerance.

At a glance

Shape
Android app (Kotlin + Flutter); on-device + cloud model hybrid
Core loop
Vision UI → tools → execution → reflection
Tools
Skills, Alpine terminal, browser, MCP, files, alarms/calendar
Memory
Short/long-term memory with embeddings
Maturity
Still prerelease as of May 2026; BestClaw watchlist
License
Open source on GitHub — see repo terms
Best for
Mobile automation early adopters accepting permission models
Risk focus
Device loss/misfire/malicious Skills; MDM and privacy compliance

Pros & cons

Pros

  • Focuses on execution loops, not a WebView chat clone.
  • On-device vision automation is a differentiated mobile experiment.
  • MCP/Skills path aligns conceptually with desktop Claw extensibility.
  • System integrations (calendar/alarms/audio) cover personal-assistant jobs.
  • Fast release cadence (80+ releases) — track now, production later.

Cons

  • BestClaw <strong>watchlist</strong> — stability, docs and enterprise cases are thin.
  • Android permission surfaces are gnarly; MDM orgs rarely approve quickly.
  • Ecosystem 5.5 — plugin/marketplace depth trails OpenClaw by orders of magnitude.
  • Prerelease builds are unsafe for core business dependencies.
  • Not directly OpenClaw-compatible — migration cost is unknown.

Capabilities (honest breakdown)

  • Vision UI automation

    Screen-driven taps; misreads cause wrong clicks — supervise risky flows.

  • Alpine terminal

    On-device shell; minimise command scope — avoid root-class exposure.

  • MCP + Skills

    Extend tools; mobile plugin story is early — audit supply chain.

  • Memory system

    Long/short context; define PII retention and wipe policies.

  • System integrations

    Alarms/calendar/audio; personal-assistant fit — enterprises need MDM.

Security — read this before go-live

Before any mobile agent go-live (especially on the watchlist):

  • Device loss — FDE + remote wipe; revoke agent credentials with the device.
  • Accessibility/screen capture — least privilege; block auto-ops on banking/OTP apps.
  • Skill supply chain — sideload audited skills only; record version hashes.
  • Hallucinated taps — human confirm transfers/deletes.
  • Enterprise MDM — IT approval for agents and model egress.

Bottom line

OpenOmniBot stays on BestClaw's mobile-agent watchlist at 6.7. Track the category; wait for stronger public signals before production. Compare PicoClaw and OpenClaw via A/B comparison and the leaderboard.

Scores and rankings follow the published BestClaw methodology; editorial and partnership placements, if any, are labeled separately and do not change numeric conclusions.

Reviews & ratings

Star ratings and review text on this page are independent of BestClaw methodology scores and leaderboard placement.

User ratings come from submissions reviewed on this page; they do not change the methodology score (6.7 / 10) or leaderboard logic.