Cline SDK v0.0.75Cline SDK v0.0.75

Cline has released version 0.0.75 of its SDK, adding provider-executed web search that lets supported models pull in live information from the internet during a single conversational turn. For developers building agentic coding workflows on top of Cline, the change matters because it shifts web retrieval from a client-managed step to something the model provider handles directly, while keeping the resulting data reproducible across sessions.

The headline addition is provider-executed web search. When a model that supports the capability is in use, it can issue search queries and receive results as part of its turn, without the host application needing to orchestrate a separate retrieval pipeline. This is distinct from client-side approaches, where the application intercepts a request, calls a search API, and feeds the results back into the prompt. Here, the search is executed by the provider's infrastructure, which typically means the model decides when a query is warranted and integrates the returned content into its reasoning.

A notable detail is how Cline handles state. The search calls and their results are persisted in session history, so they replay when a session is reloaded. This is important for reproducibility and debugging: an agent's behavior often depends on the exact information it retrieved at a given moment, and without persistence, reloading a conversation could produce different results or lose the context that shaped earlier responses. By storing both the calls and their outputs, Cline preserves a faithful record of what the model saw.

The feature is off by default. Users who want it must enable the web_search model tool in settings, making this an opt-in capability rather than a behavior that changes existing workflows automatically. That default is a reasonable choice given that web search introduces external data, potential latency, and, depending on the provider, additional cost per query. Teams that need deterministic or offline-friendly behavior can leave it disabled.

The release notes also indicate the addition of a dedicated Cline provider, though the available excerpt is truncated. Based on the wording, it appears Cline is introducing a first-party provider option alongside its existing support for third-party model backends, though the specifics are not fully described here.

Some background helps frame the update. Cline is an open-source AI coding assistant that operates as an agent, reading and editing files, running commands, and iterating on tasks within a developer's environment. It is model-agnostic, working across providers such as Anthropic, OpenAI, and others, and the SDK exposes this functionality programmatically so it can be embedded into other tools and automated pipelines. An SDK release like this one is aimed less at end users clicking through an editor and more at developers wiring Cline into their own systems.

Provider-executed web search itself reflects a broader industry pattern. Major model vendors have increasingly built server-side tools, including web search, into their APIs, so that the model can call these capabilities natively rather than relying on the application to simulate them. Anthropic and OpenAI, among others, have shipped hosted tool-use features along these lines. Cline's decision to surface provider-executed search, and to persist its artifacts, aligns the SDK with that direction while addressing a practical gap around session continuity.

For practitioners, the immediate takeaways are straightforward. The capability is limited to models that support it, so the benefit depends on the chosen backend. Because results are stored in session history, reloaded conversations should behave consistently, which is useful for audit trails and for reproducing agent runs. And because it is opt-in, existing deployments are unaffected unless the web_search tool is explicitly enabled.

As with any point release in a fast-moving project, the full impact will likely become clearer as more providers expose compatible search tooling and as developers report how persistence behaves in longer-running sessions. For now, v0.0.75 is best read as an incremental but meaningful step toward richer, reproducible agent behavior.