Stop Tapping Tools One Move at a Time—OpenCode 2.0 Folds Away Terminal Wait Time with Background Sidekicks and "Script Execution"

OpenCode, the open-source coding agent that lives in your terminal, has officially shipped its major 2.0 release. It's built by SST (Serverless Stack), the license remains MIT, and the tool itself is free. Press coverage has described it as "a major overhaul." This stable release follows a public beta that began about three months ago, and the headline features come down to three things: a shared background service, background agents, and Code Mode.

Three front ends now look at one resident process

Until now, OpenCode's terminal UI, web app, and desktop app each ran on their own. In 2.0, they've been reworked to connect to a single background service. The server launches automatically as one resident process, and the TUI, web, and desktop apps all talk to that same process.

As a result, sessions are no longer split per front end—they're shared across them. Picking up in the browser right where you left off in the terminal is now the default way to work. On top of that, the @opencode/client package auto-detects the local service, so scheduled scripts like cron jobs run by simply "connecting." It's a design built with unattended operation and scheduled execution in mind.

"Waiting" tasks can now be handed off to the background

In 2.0, you can run subagents and shell commands asynchronously. The idea is that you can kick off a test suite run or a longer investigation in the background while the main agent keeps working on something else. A task that started in the foreground can also be pushed to the background mid-run by pressing Ctrl+B.

For example, you can hand a code review off to a subagent running on a different AI model while you push ahead with the core implementation. When the heavy job finishes, the main agent gets a notification, so you spend less time idly waiting for completion. It's a shift away from "a human babysitting every single move" toward "fire it off and catch the result."

Code Mode—bundling one-at-a-time "check-ins" into a single script

The most novel addition this time is Code Mode. It's a small JavaScript interpreter: instead of the agent "calling one tool per turn," it executes a short script that chains multiple operations together in a single round trip. MCP servers and OpenAPI endpoints are treated as functions the script can call, and all MCP servers are pulled in by default.

Running the chain of "call a tool, transform its result, call another tool" as one script pays off in two ways:

• Fewer round trips: You cut the back-and-forth of querying the model at each sequential step, which shrinks the number of moves itself.
• Context savings: Intermediate data generated along the way never enters the model's context, so you curb context consumption.

That said, the makers themselves add a caveat about these context savings: they "haven't run 2.0 themselves yet and haven't seen independent benchmarks." Hard numbers to back it up are still to come.

A face as an "embeddable development platform"

For developers, there's a new embeddable @opencode/sdk that emits the same code as the CLI. You can also strip the build down for servers or Durable Objects. The API has been reworked to a standard server protocol, in a setup where OpenCode calls its own API (for git worktree operations, session migration, and the like). The plugin system has been expanded too: TUI plugins are now stable and support hot reload.

On the business side, its model-agnostic character carries weight. OpenCode supports a wide range of providers—from Claude, OpenAI, and Gemini to local LLMs—the tool is free, and you only pay for the usage of whichever model you choose. As an option that doesn't lock you into a specific vendor's monthly subscription and lets you swap models to your own policy, it fits comfortably into in-house or on-prem-oriented shops.

Behind the convenience, some things get harder to inspect

At the same time, a design that dials up autonomy calls for caution. The more subagents running in the background, the harder it becomes for a person to grasp "what's running right now," and the more you run things unattended, the heavier the burden of monitoring and cost management is likely to get. Code Mode, too, is worth keeping in mind: since the agent generates and runs scripts, it carries the new dimension of executing generated code. These concerns aren't unique to OpenCode—they're the kind of thing pointed out about self-driving agents in general.

There are practical incompatibilities as well. 2.0 isn't compatible with 1.0 plugins, and the server API has been redesigned. Users with existing extensions or scripts will need to do some work to migrate. In exchange for the convenience, you should expect some rebuild cost at the adoption stage.

References: The New Stack: OpenCode 2.0 puts its terminal, desktop, and web apps on one shared service / AlternativeTo: OpenCode 2.0 officially launches with background agents, Code Mode, and a rebuilt API / ExplainX: OpenCode 2.0 — Code Mode, Background Agents, SDK Upgrade Guide