OpenCode v2.2.6: Shared MCP, worktrees and archived runs

Anyone running OpenCode across several repositories at the same time feels the problem: each project opens its own connection to the same MCP server, consumes more processes and takes longer to start. v2.2.6 fixes this with connection pooling and adds two new commands that help with automation and CI.

In this article you will see:

• How shared MCP works across projects
• How to run a session in a new worktree with opencode run -w
• How to archive CI runs with the --archived flag
• Other release improvements and how to test today

What changed in OpenCode v2.2.6

OpenCode 2.2.6 brings five main highlights. The first is sharing MCP connections between projects that use the same configuration. The second is the ability for UI plugins to add color themes, actions in code blocks and native configuration components. The third is a composer dock for phones and tablets. The fourth is running runs in git worktrees. The fifth is automatic archiving of CI runs.

Beyond the highlights, the release brings core improvements, such as session warming by provider and model, a usage policy for web search, an experimental bytecode VM for Code Mode and support for PowerShell 7 installations from the Microsoft Store on Windows.

Shared MCP across projects

Sharing MCP connections is the most useful change for people who maintain several projects. Until now, each project opened its own connection to the same MCP server. With v2.2.6, just enable pooling with pool: true in the server configuration. All projects with the same configuration will start using a single connection.

The practical result is fewer processes running at the same time and faster startup. For people who use heavy MCP servers, such as search or file servers, the resource savings are noticeable.

To enable it, add pool: true to the MCP server entry in the OpenCode configuration file. Then run OpenCode in two different projects and check that only one server process appears. If the server requires per project state, disable pooling in that case.

Runs in worktree with opencode run -w

The flag --worktree, or -w, starts a session in a new git worktree. This is useful when you want to test a change without touching the main project tree. Just run opencode run -w and OpenCode creates the worktree and starts the session in it.

If you want to choose the worktree name, use --worktree=. The flag cannot be combined with --continue or --session, because the new session has no previous history. People who use worktrees in CI can create a worktree per run and discard it at the end.

A common flow is to run opencode run -w "crie testes para o módulo X" and let the agent work in the isolated worktree. If the result is good, you merge the branch. If not, you discard the worktree without dirtying the main project.

Archived sessions for CI runs

A flag --archived creates the session already archived. CI runs and automated scripts stop appearing in the inbox and session lists. Archived sessions are also hidden in the terminal interface and are ignored by --continue.

This solves a common problem: after running dozens of agents in CI, the session list gets filled with runs that no one will review. With --archived, those runs stay out of the way, but can still be consulted when needed.

To use it, add --archived to the opencode run command. In pipelines, combine with --format json to capture output in a structured way. Remember to set timeouts on CI jobs, because OpenCode now waits for background commands, subagents and image jobs to finish before exiting.

Other improvements worth testing

The release also changes the behavior of opencode run in non-interactive mode. Now the command waits for background commands, subagents and image jobs started by the prompt and lets the model respond to the results before exiting. While waiting, it prints waiting for N background job(s). There is no time limit: a job that never ends, like a development server, keeps the run open until you press Ctrl+C or send SIGTERM. In CI, set a timeout.

SIGTERM handling also changed. Now the signal cancels the session and background jobs and exits with code 143. This makes OpenCode more predictable in pipelines that need to interrupt executions.

There is also session warming per provider and model, with interval: "auto" mode based on cache TTL. And the websearch.use policy to control web search usage. If you want to test experimental features, you can enable all flags with OPENCODE_FLAGS=all or experimental.all_flags.

How to test today

1. Update OpenCode to version 2.2.6.
2. In two projects with the same MCP server, add pool: true and check the process reduction.
3. Run opencode run -w "descreva a tarefa" in a git repository and verify the worktree created.
4. Run opencode run --archived "tarefa de CI" and check that the session does not appear in the main list.
5. In a pipeline, add timeout and test the behavior with SIGTERM.

If you already use OpenCode with other agents, it's worth comparing the flow with the MCP server git and with the Claude Code 2.1.319, which also changed MCP and background.