Cline open-sources its agent SDK: the CLI already runs on it, VS Code and JetBrains are next
The Cline team has extracted the AI agent's engine from its VS Code extension into the open-source @cline/sdk package. Previously, the agent loop grew inside the IDE plugin, which made it hard to maintain and to embed in other products. The CLI now runs on the SDK, and the VS Code and JetBrains extensions are migrating; the package includes subagents, plugins, cron, and MCP. Teams can take a ready-made foundation for their own coding agent instead of writing orchestration from scratch.
AI-processed from Cline Blog; edited by Hamidun News
Cline was one of the first to show what agentic coding looks like in an editor: back on Sonnet 3.5, it taught the model to call tools, before Claude Code and Codex appeared. Now the team acknowledges that the architecture that let it reach the market early is poorly suited to the next stage. So the agent engine has been moved out of the extension into a separate open-source package, @cline/sdk. It can be installed with npm i -g @cline/sdk and embedded in your own product.
Why the foundation had to be rewritten
The story is typical for early agent products. First there was a VS Code extension, and the agent loop grew directly inside it. Then terminal integration, computer control, subagents, multi-agent teams, and custom tools were added. The product became smarter, but the engine grew ever more tightly bound to the IDE. It was hard to maintain, extend, and reuse.
Now the runtime (the agent's execution environment) has been turned into a shared service rather than a detail of one application. The low-level agent loop stays stateless and reusable. Around it sits a durable layer that stores state. Practical consequences follow: a long task doesn't die when the interface restarts, and a session can be moved from one client to another.
Cline 2.0 is a layered TypeScript stack. Each layer handles one job and depends only on the layer below it. If you need the whole set, you take it all. If you need only the provider layer, for example for an LLM proxy, or only the stateless loop for a serverless environment, you take individual packages. You connect state when you actually need it.
What's in the box
Through plugins, teams add their own logic without forking the runtime. A plugin can register tools, watch lifecycle events, add rules and commands, and change what the agent sees. You can start with a local .ts or .js file. Once the solution matures, you package it with a cline.plugins manifest. There are other extension points too: custom tools, MCP, and skills.
Providers connect through a single layer. Supported are Anthropic, OpenAI, Google, AWS Bedrock, Mistral, LiteLLM, and OpenAI-compatible endpoints such as vLLM, Together, and Fireworks. According to the authors, switching providers comes down to editing a config. Adding your own provider takes a call to llms.registerHandler("my-provider", handler).
Subagents and agent teams are built into the runtime. A session can delegate parts of the work to specialists, each with its own model, tools, and prompts. They exchange statuses and notes when handing off tasks. You don't need to write a separate orchestration layer. In addition, cron jobs, checkpoints, web search, and MCP connectors work out of the box.
What it means in practice
Not the entire lineup has moved yet. The CLI and kanban have already been ported to the new SDK, while the VS Code and JetBrains extensions are still migrating. Cline says its applications are used by more than 7 million developers. The harness itself has also been rewritten: the prompts, a simplified loop, more careful context management, error handling, and tool descriptions. These improvements live in the runtime, so all clients get them at once.
The team reports results for Cline CLI on terminal-bench 2.0 (the pass@1 metric, the share of tasks solved on the first attempt) and compares it with OpenCode and Pi-Code on open-weights models. These comparative runs were carried out by the Cline team itself on versions as of May 8, 2026, so the data is not independent. On speed and token usage, the material only says that in internal runs the new CLI handles the same tasks faster and more cheaply than the old one. No specific figures are given in the article.
The new CLI has gained experimental channel connectors for Telegram, WhatsApp, and Slack. They are set up through an interactive wizard. The authors suggest these scenarios: a support bot that reads the codebase before answering, a nightly dependency-audit agent that opens a PR and assigns reviewers, and a multi-agent workflow that turns a one-line bug report into a tested fix.
For developers and product managers, the takeaway is simple. If you were building your own agent from scratch or forking someone else's IDE plugin, there is now an open foundation with a ready-made loop, providers, and subagents. Before building production on it, check the stability of the API: the extension migration is still underway, and the project asks users to try the examples and report what breaks.
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
Why the foundation had to be rewritten?
The story is typical for early agent products. First there was a VS Code extension, and the agent loop grew directly inside it. Then terminal integration, computer control, subagents, multi-agent teams, and custom tools were added. The product became smarter, but the engine grew ever more tightly bound to the IDE. It was hard to maintain, extend, and reuse. Now the runtime (the agent's execution environment) has been turned into a shared service rather than a detail of one application.
What's in the box?
Through plugins, teams add their own logic without forking the runtime. A plugin can register tools, watch lifecycle events, add rules and commands, and change what the agent sees. You can start with a local .ts or .js file. Once the solution matures, you package it with a cline.plugins manifest. There are other extension points too: custom tools, MCP, and skills. Providers connect through a single layer.
What it means in practice?
Not the entire lineup has moved yet. The CLI and kanban have already been ported to the new SDK, while the VS Code and JetBrains extensions are still migrating. Cline says its applications are used by more than 7 million developers. The harness itself has also been rewritten: the prompts, a simplified loop, more careful context management, error handling, and tool descriptions. These improvements live in the runtime, so all clients get them at once.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.