opencode-claude-code-bridge

opencode-claude-code-bridge

OpenCode plugin that bridges Claude Code's MCP server configs and the skills bundled inside Claude Code plugins into OpenCode — so switching between the two CLIs (or running them side-by-side) doesn't require re-wiring anything.

It also brings Claude Code's skills-as-slash-commands ergonomics to the OpenCode TUI: every discovered skill becomes a first-class /<skill> command you can type and autocomplete, just like in Claude Code.

What it pulls in

• User-level MCP servers from ~/.claude.json (mcpServers).
• Project-level MCP servers from <cwd>/.mcp.json. If .claude/settings.json explicitly declares enabledMcpjsonServers, that allowlist is respected; otherwise the local file is imported as-is.
• Plugin-bundled MCP servers from any enabled Claude Code plugin (~/.claude/plugins/installed_plugins.json ∩ enabledPlugins).
• Plugin-bundled skills — symlinked from ~/.claude/plugins/cache/<marketplace>/<plugin>/<version>/skills/<name> into ~/.config/opencode/skills/<name> (OpenCode already discovers skills at that location).
• Skills as TUI slash commands — every discovered skill is surfaced as a first-class /<skill> slash command in the OpenCode TUI, for parity with Claude Code (see below).

For MCP definitions that reference ${CLAUDE_PROJECT_DIR}, the bridge resolves it to OpenCode's current project directory.

Not yet supported (v1): bundled agents, hooks.

Skills as slash commands (TUI parity with Claude Code)

In Claude Code, every skill is also a typeable /<skill> slash command. OpenCode treats skills differently per client:

• Desktop already renders skills as / slash entries natively.
• TUI does not — its slash palette only lists source: "command" entries, so out of the box skills are reachable in the TUI only via the /skills picker, never as /<skill>.

This plugin closes that gap. On non-Desktop clients (TUI, CLI run, headless) it injects a source: "command" wrapper for each discovered skill, so /share, /list-repos, etc. autocomplete in the TUI slash menu exactly like in Claude Code.

Key properties:

• No duplicates in Desktop. The plugin runs inside each client's own server process and gates on OPENCODE_CLIENT; under Desktop (OPENCODE_CLIENT=desktop) it injects nothing and leaves Desktop's native skill rendering untouched.
• No transcript spam. The wrapper is a one-line shim — it tells the agent to invoke the skill via its skill tool and forwards your $ARGUMENTS. The skill body loads as the skill tool result, so /share doesn't dump the whole SKILL.md into your message. Typing /list-repos --limit 20 forwards --limit 20 to the skill.
• Covers every skill root, not just plugin-bundled ones: ~/.config/opencode/skills (global + bridged), ~/.claude/skills (Claude Code "external" skills, e.g. those installed by other tooling), and ~/.agents/skills.
• User config always wins. A command you define yourself (or an OpenCode built-in) with the same name is never overwritten.

This is config-hook injection only — no files are written to your ~/.config/opencode/command/ directory, so nothing leaks across clients or persists on disk.

Install

Add the plugin to your OpenCode config (~/.config/opencode/opencode.jsonc or project-level .opencode/opencode.jsonc):

{
  "plugin": ["opencode-claude-code-bridge@latest"]
}

If OpenCode already has an authoritative MCP catalog and you only want Claude plugin skills and slash-command parity, disable Claude MCP discovery:

{
  "plugin": [["opencode-claude-code-bridge@latest", { "mcp": false }]]
}

This leaves explicit OpenCode MCP entries untouched and skips user-level, project-level, and Claude-plugin-bundled MCP imports. Skill bridging remains enabled. MCP import defaults to true for backward compatibility.

OpenCode will install it from npm on next start. The @latest tag ensures you automatically get new versions on each restart (ref). To pin a specific version instead, use e.g. "opencode-claude-code-bridge@0.1.1".

When does OpenCode pick up Claude Code changes?

On OpenCode restart. OpenCode loads its config once at startup and does not hot-reload, so the same applies here — every time you start OpenCode, this plugin re-reads Claude's configs from scratch and injects the current state. No manual sync command needed; just quit and relaunch OpenCode.

What that means in practice:

You change in Claude Code…	Visible in OpenCode after…
Add/remove an MCP server in ~/.claude.json or .mcp.json	OpenCode restart
Toggle a plugin in enabledPlugins	OpenCode restart
Edit enabledMcpjsonServers	OpenCode restart
Install a new Claude plugin (new entry in installed_plugins.json)	OpenCode restart
Add/remove a skill (new /<skill> TUI slash command)	OpenCode restart
Edit the contents of a skill file (e.g. SKILL.md body)	Immediately — skills are symlinked, not copied

Conflict resolution

If you have an MCP server name declared in both your opencode.jsonc and Claude's configs, your opencode.jsonc entry wins.

Debugging

Set OPENCODE_CLAUDE_CODE_BRIDGE_DEBUG=1 before launching OpenCode to log which MCPs, plugins, and skill command wrappers were discovered (the latter logs as injected skill command wrappers: [...], or desktop client — skipping… under Desktop).

To confirm a skill is now a slash command in the TUI, check that it appears with source: "command":

opencode debug skill   # lists discovered skills
# In the TUI, type `/` and your skill names should autocomplete.

Publishing

This package publishes through GitHub Actions trusted publishing from .github/workflows/publish.yml.

For maintainer release steps and verification, use:

• .claude/skills/publish-package/SKILL.md