The open-source personal agent platform OpenClaw just dropped version v2026.5.18, and it's a big one. The Android client now swaps text-based chat for real-time voice conversations relayed through a gateway, and the update fully opens the floodgates for GPT-5 series models.
On Android, the new client handles streaming microphone input and real-time audio playback, and it bridges tool results — syncing tool call status with the voice stream in real time — along with live screen captions. That means you can just talk to wake up and run complex local toolchains from your phone.
On the model front, OpenClaw has removed the configuration validation blocks for GPT-5.1, GPT-5.2, GPT-5.3, and the openai-codex model. It also kills the old logic that forcibly truncated GPT-5's final responses, so you get the full channel output now. When strict-agentic execution is enabled, the system automatically writes logs.
Plug-in development gets simpler with the new defineToolPlugin minimal interface and companion CLI tools: openclaw plugins build, validate, and init. Developers can declare simple tool plugins with strong typing, and the system auto-generates the manifest and context factory.
Under the hood, the Memory-core introduces incremental sync on startup. It compares disk sessions against index files and only re-indexes missing, changed, or resized files, drastically cutting cold start times.
Finally, untrusted signal configuration reloads — like SIGUSR1 — now happen entirely in-process, avoiding the PID loss that came from forking orphan children.