openclaw 2026.3.24-beta.1

OpenClaw v2026.3.24-beta.1 arrived this week with a change that does not announce itself loudly but signals something real: the platform now forwards /v1/models and /v1/embeddings requests alongside its existing /v1/chat/completions and /v1/responses endpoints, and passes explicit model overrides through to both. In plainer terms, OpenClaw is no longer just capable of calling OpenAI models — it is now targetable by the same client SDKs, RAG pipelines, and agentic workflows that assume an OpenAI-shaped interface exists at the other end of the wire.

That distinction matters. Client libraries for agents and retrieval-augmented generation pipelines do not arbitrate between providers at runtime — they send requests to a fixed endpoint structure and expect model discovery to work a certain way. When a RAG pipeline hits /v1/models, it wants a list of available models in OpenAI's format. When a multi-model agent passes an explicit model parameter, it expects that parameter honored, not silently dropped. Adding those endpoints sounds routine. It is, in fact, the thing that moves OpenClaw from "works with OpenAI models" to "works with OpenAI-compatible tooling" — and in an ecosystem where agentic applications are increasingly built against a OpenAI-shaped contract, that is a meaningful shift toward interoperability as a first-class design goal rather than a fortunate side effect.

The release, posted to GitHub, includes several other changes worth examining on their own terms rather than as features in a changelog.

The before_dispatch plugin hook — contributed by developer @gfzhx in pull request #50444 — fires after routing, send policy, and command parsing have all run, but before the model itself is called. The plugin can then short-circuit with a handled reply, bypassing the LLM entirely for cases where the infrastructure already knows the answer. This is not a dramatic rearchitecture. It is a single call site in dispatchReplyFromConfig(). But for developers building plugins that need to intercept or redirect requests at the last moment — before a round-trip to a remote model — it closes a gap that required awkward workarounds before.

Microsoft Teams gets a migration to the official Teams SDK, which carries its own set of AI-agent UX improvements. The release notes enumerate them with unusual specificity for a changelog: streaming one-to-one replies, welcome cards with prompt starters, feedback and reflection flows, informative status updates, typing indicators, and native AI labeling. These are not revolutionary features. They are the specific UX primitives that make an agent feel like an agent rather than a bot that prints one message and goes silent — and shipping them as documented best practices alongside the SDK migration means teams adopting the new integration do not have to rediscover them from first principles.

The CLI picks up a --container flag and a corresponding OPENCLAW_CONTAINER environment variable, enabling docker exec-style execution of OpenClaw commands inside a running container. This is infrastructure for infrastructure: if you are already running OpenClaw inside Docker or Podman and want to inspect, configure, or trigger a skill from the host without attaching a shell, the flag exists. It is a small ergonomic improvement that reflects the growing reality of OpenClaw deployments in containerized environments.

Skills metadata gets a one-click install recipe system for bundled skills — coding-agent, gh-issues, openai-whisper-api, session-logs, tmux, trello, and weather. The CLI and Control UI can now detect missing dependencies when a skill is invoked and offer to install them rather than failing opaque. Again, small. But the gap between "install the skill" and "install the skill and all the things it needs to actually run" is where many developers lose an afternoon, and automating that detection is the kind of plumbing that makes the difference between a skill being usable and being a documentation exercise.

One addition that deserves explicit mention: a security fix closing the mediaUrl/fileUrl alias bypass in sandboxed media dispatch, documented as pull request #54034. The bypass allowed outbound tool and message actions to escape media-root restrictions under certain aliasing conditions. No specifics on active exploitation are noted in the release; the fix resolves the vector regardless. Agent platforms handling user-provided content should treat this as a routine-but-urgent patch.

What the release does not say is as instructive as what it does. The OpenAI compatibility layer is the strategic headline even if it reads like a technical footnote. OpenClaw's primary product is not a model — it is an agent execution environment. Making that environment reachable by tooling that expects an OpenAI API surface is a bet that the OpenAI Responses API is becoming the de facto agent-client interface, and that platforms which speak that language fluently will attract developers who have already built against it. That is a coherent hypothesis about where the ecosystem is heading. Whether it holds depends on whether the Responses API standard actually settles — and whether OpenClaw's execution model is differentiated enough to justify running on top of it rather than alongside it.

OpenClaw v2026.3.24-beta.1 is available on GitHub. The changelog is clean, the security fix is credited, and the new endpoints are documented with enough specificity to write against. Worth watching: whether subsequent releases extend the model override passthrough to more routes, and whether the before_dispatch hook develops a library of reusable plugins that take advantage of it.