OpenClaw v2026.4.25 Ships Full TTS Upgrade With Azure Speech and OpenTelemetry

OpenClaw v2026.4.25 ships a broad text-to-speech update that adds Azure Speech as a bundled provider, expands voice control at the agent level, and extends observability across the platform with OpenTelemetry. The release landed on April 25 and is listed in the project’s release notes on GitHub.

The headline change is the TTS overhaul. OpenClaw now lets users set per-agent voice overrides, which means different agents can speak with different voices instead of sharing one global setting. That matters for teams building multi-agent workflows, where a support agent, a scheduling agent, and a research agent may need distinct voices or presentation styles.

Azure Speech joins the built-in provider list alongside Xiaomi, Inworld, Volcengine, and ElevenLabs v3. By bundling Azure rather than treating it as an add-on, OpenClaw makes it part of the core release path for users already running Microsoft cloud services. For operators, that can simplify setup and reduce the amount of glue code needed to wire speech output into an agent pipeline.

The release also broadens OpenTelemetry coverage. OpenTelemetry is an open standard for collecting telemetry data, such as traces, metrics, and logs, from applications. In practical terms, that gives operators better visibility into where requests are going, how long they take, and what happens inside the system as agents call models and tools.

According to the release notes, the new telemetry coverage includes model calls, token usage, and tool loops. Those are three of the most useful places to instrument an AI automation stack. Model calls show when the system is querying an LLM or other model, token usage helps track consumption and cost, and tool loops show repeated back-and-forth between an agent and external functions such as search, databases, or APIs.

That kind of visibility is especially valuable in systems where one request can fan out into multiple steps. If an agent is calling tools repeatedly, or spending more tokens than expected, those patterns are much easier to spot when the telemetry is built into the platform instead of stitched together later.

OpenClaw v2026.4.25 also changes how the plugin registry is stored. The release moves it to a cold persisted path, which means the registry lives in storage that is kept across restarts rather than being rebuilt from scratch each time. The practical effect is faster startup, especially for installations with a larger plugin catalog.

For self-hosted deployments, startup time can matter more than it does in a small local setup. A plugin registry that has to be recreated repeatedly slows down restarts, complicates updates, and makes failures more annoying to recover from. Persisting that state reduces that overhead and keeps the system closer to ready when the service comes back up.

The release fits into a broader pattern in OpenClaw’s development: more native support for speech, better instrumentation for AI workflows, and less startup friction for plugin-heavy installs. Those are the kinds of changes that tend to matter most once teams move from demos to production-style automation, where voice output, tracing, and restart behavior all become part of day-to-day operations.