Changelog

Each release is tagged with a version number following semantic versioning: major.minor.patch. A major version bump indicates breaking changes that require migration steps. A minor version bump indicates new features that are backwards compatible. A patch version bump indicates bug fixes with no behavioral changes.

We announce breaking changes at least one release in advance. When a major version ships with breaking changes, the previous minor version remains supported for 60 days to give you time to migrate. During this window, both versions receive security patches.

Subscribe to the releases feed in your dashboard to receive email notifications when new versions ship. For hosted customers, we handle runtime updates automatically during your next maintenance window. For self-hosted and local deployments, you control the update timing.

Runtime 3.0 is the most significant release since OpenClaw 1.0. The streaming task output feature allows agents to stream incremental results for tasks that produce large outputs, rather than waiting for the complete result before returning anything. This improves responsiveness for long-running data processing and code generation tasks.

The v2 skills manifest introduces a new dependency resolution system that eliminates a class of dependency conflicts between skills. Existing skills using the v1 manifest continue to work with the v1 compatibility shim, but we recommend migrating to v2 at your next convenience. The migration tool converts v1 manifests to v2 automatically in most cases.

The relay disconnection fix addresses a long-standing issue where agents running more than 8 hours under continuous load had an increasing probability of WebSocket connection instability. The 40% reduction in disconnection rate was achieved through changes to the heartbeat protocol and connection pool management.

The Model Switch skill was one of the most requested features from enterprise customers. It provides a provider-agnostic interface for switching between OpenAI, Anthropic, and local model endpoints without modifying agent logic. Configure the active provider in your dashboard or via the CLI, and all agents using the Model Switch skill immediately use the new provider.

Browser Relay skill v2 adds native Safari support on macOS 14 and later, in addition to existing Chrome support. Safari support uses the same WebSocket relay infrastructure as Chrome, so relay server configuration does not need to change. Install the Safari extension from the OpenClaw dashboard under Settings > Browser Extensions.

The CLI skills commands (install, update, validate) give skill developers a faster workflow for managing skill lifecycles. The validate command runs a full dependency check and smoke test before a skill is activated, preventing broken skills from entering your active skill set.

The multi-agent fleet dashboard was the flagship feature of the 2.5 release. For the first time, teams running more than one agent have a unified view of fleet health, task queue depth, and per-agent resource utilization. The dashboard updates in real time via WebSocket and supports fleet-wide filtering and search.

Webhook delivery with automatic retry addresses a common integration requirement: reliable event notification even when your webhook endpoint is temporarily unavailable. Failed deliveries are retried with exponential backoff (5s, 30s, 1m, 5m, 30m, 2h) for up to 6 attempts before being marked as permanently failed.

The published OpenAPI spec enables teams to generate type-safe API clients for their language of choice using standard OpenAPI tooling. This makes it practical to integrate ClawMesh agent capabilities into existing applications without using one of the official SDKs.

Docker Compose template improvements in 2.4 added GPU passthrough for NVIDIA GPUs on Linux hosts, enabling GPU-accelerated skills for the first time in containerized deployments. The compose.yaml template was updated with the necessary device mappings and runtime arguments.

The Skills Hub reached 47 managed skill sets covering domains including code analysis, data processing, document generation, API integration, and media handling. All managed skills are tested as complete units and updated in coordination, eliminating the version management overhead that previously burdened custom skill users.

The macOS one-click installer for Apple Silicon resolved the most common setup failure point for Mac users: Node.js version mismatches. The installer bundles a verified Node.js runtime and does not depend on any system-level Node.js installation.

Runtime 2.3 introduced multi-agent support, allowing a single workspace to run more than one agent concurrently. The fleet management CLI (clawmesh fleet) was added for managing agent lifecycle from the command line. This laid the groundwork for the fleet dashboard that shipped in 2.5.

Runtime 2.2 shipped the Telegram bot integration that enables mobile access to agents without requiring the dashboard. The bot forwards messages to the agent runtime and sends responses back as Telegram messages. This was the most impactful quality-of-life improvement for users who wanted agent access from mobile devices.

Runtime 2.1 shipped the first version of the browser relay and Chrome extension. The relay enables agents to interact with web applications programmatically by injecting commands into browser sessions. This opened up automation use cases that require interaction with web UIs rather than APIs.