OpenCode 2.0 Rewrites Its Core: MCP Tools Can Be Chained with Code, Multiple Agents Can Work in Parallel in the Background

Beating AI News Flash: Open-source AI coding tool OpenCode 2.0 has officially launched, following its public beta in July. A major change in this upgrade is Code Mode, which can convert MCP services and OpenAPI interfaces into JavaScript functions, allowing Agents to chain multiple tools within a single script. For example, it can first batch-read bugs from GitHub, then create Linear tasks one by one, without requiring the model to re-initiate calls for every piece of data processed. All MCP services enable Code Mode by default, which can also reduce the context occupied by tool definitions.

OpenCode has also rewritten its underlying architecture. The terminal, web, and desktop clients now share the same local backend service and can access the same set of sessions. Sub-agents and Shell commands can also run in the background, for example, letting one Agent review code while the main Agent continues modifying files and simultaneously runs tests and builds. The team has also turned a large number of internal features into plugins, now exceeding 80, supporting runtime hot-loading, so Agents can write plugins for themselves and use them immediately.

In addition, OpenCode has launched an SDK that can be embedded into other applications. Cloudflare has announced experimental interface support for OpenCode 2.0, allowing developers to run the same Agent engine in Durable Objects and save sessions and resume interrupted tasks.

However, this refactoring has also brought compatibility issues. 1.x plugins need to be rewritten, and the old Server API also needs adaptation. An issue on GitHub points out that Code Mode lacks a timeout limit when calling long-running plugins, which may leave sessions waiting for a long time and require manual interruption.