opencode-subagent-mcp
Run Claude Code subagents on cheaper opencode models — for example
opencode-go/deepseek-v4.1-flash (DeepSeek) from an opencode Go subscription — while Claude
Code itself keeps talking to Anthropic directly. Delegate searches, refactors, bug fixes, tests
and research to a model that costs a fraction of Opus or Sonnet, and keep the tools you already
use: background agents, resume via send_message, worktree isolation, .claude/agents
definitions, permission prompts and structured output.
It is a stdio MCP server. Your main Claude Code session is not proxied or routed
anywhere; the server only runs the subagents you delegate to it, on a private
opencode serve instance it starts itself.
Install
Requires Node 22+, git, and the opencode CLI logged in to the
providers you want (opencode auth login).
As a Claude Code plugin (also adds a skill that tells Claude when to delegate to opencode):
Or as a plain MCP server:
What it runs and sends
- It starts a private
opencode serve process on a random localhost port with a random
password, and stops it when Claude Code exits. The password is generated for each run and
only used between this server and that local process. The process inherits your environment
so opencode can use the provider logins and API keys you already configured for it; this
server never reads, stores or sends those credentials itself.
- The prompts you delegate, and the files and command output the subagent reads, go to the
model provider you choose through opencode (by default opencode Go), under that provider's
data retention terms. opencode keeps session history in its local database on your machine. Nothing else is sent
anywhere; the server has no telemetry and makes no other network calls.
- It reads
~/.claude/settings.json, the project's .claude/settings.json and
.claude/settings.local.json, and .claude/agents/*.md (user and project) to mirror your
permission rules and agent definitions, and
~/.config/opencode-subagent-mcp/config.json if present.
- With
isolation: "worktree" it creates git worktrees under the project's
.claude/worktrees/ and removes them when they have no changes.
This is an independent project, not affiliated with Anthropic or the opencode team.
Tools
An agent_id keeps working after the MCP server or Claude Code restarts: wait and stop
return what happened to it, and send_message continues it. Agents from before a restart do not
appear in list until their id is used again, but they still exist in opencode's local session
store.
agent parameters: description, prompt, subagent_type, model, isolation: "worktree",
permission_mode, tools, disallowed_tools, max_turns, effort, output_schema,
run_in_background, cwd.
Parity with native subagents
Known gaps: agents do not appear in Claude Code's /tasks agent view; there is no
fork-with-parent-context; the Workflow tool's agent() cannot target them; auto mode has no
safety classifier (it allows everything, including paths outside the project, except rules
you configured and a fixed list of destructive commands that ask: rm -r/-rf, git push --force, git reset --hard, git clean -f, git restore/git checkout --, curl | sh,
sudo, chmod -R/chown -R, dd, mkfs, npm/pnpm publish; opencode matches each
command in a chain separately); the parent's permission mode is not visible to MCP servers, so
the default comes from config; hooks, skills preloading, memory and per-agent MCP servers are not mapped.
Steering Claude
Claude Code picks between its native Agent tool and this one on its own, so routing is
unpredictable. To make opencode the default for delegated work that is cheap to verify, paste the snippet from docs/claude-md-snippet.md into
your global ~/.claude/CLAUDE.md.
Configuration
Optional ~/.config/opencode-subagent-mcp/config.json (or $OPENCODE_SUBAGENT_CONFIG):
OPENCODE_SUBAGENT_MODEL and OPENCODE_SUBAGENT_MODE override the first two. pure starts
opencode without external plugins.
How it works
On first use the server starts opencode serve on a random loopback port with a random
password, subscribes to its event stream, and drives one opencode session per agent
(agent id = session id). Permission requests from opencode are answered through MCP
elicitation. The opencode process is stopped with the MCP server; leftovers from crashed
servers are reaped on the next start.
Development
License
MIT
