OpenCode

Running opencode starts both a server and a TUI client that talks to it; the server publishes an OpenAPI 3.1 spec, which is also used to generate the JS/TS SDK. opencode serve runs the server alone, so scripts, the desktop app, IDE extensions or opencode run --attach <url> can drive the same sessions.

Multi-agent work follows a supervisor model inside one session tree. A primary agent (Build, Plan or one you define) can call subagents through the Task tool, either on its own based on each subagent's description or when the user @-mentions one. Each subagent call creates a child session that the user can open, cycle through and return from. Subagents inherit the calling agent's model unless configured otherwise, and permission.task controls which subagents each agent may invoke.

Project instructions live in an AGENTS.md file (created with /init), which is loaded into context. Long sessions are shortened by a hidden compaction agent, configurable through the compaction option; there is no separate long-term memory store documented.

Every tool call is checked against the permission config: allow runs it, ask prompts the user, deny blocks it. Rules can target patterns (for example "git *": "allow", "rm *": "deny") and differ per agent; the built-in Plan agent asks before any edit or bash command. --auto (or the TUI command palette) approves everything not explicitly denied. The question tool lets the model ask the user multiple-choice or free-text questions mid-task. Users can open any subagent's child session to watch it, and can @-mention a subagent directly even when an agent's task permissions deny it.

npm install -g opencode-ai

npm install -g opencode-ai
opencode auth login            # store an API key for your provider

# a project-level review subagent that cannot edit files
mkdir -p .opencode/agents
cat > .opencode/agents/reviewer.md <<'EOF'
---
description: Reviews diffs for bugs and risky changes
mode: subagent
permission:
  edit: deny
---
List concrete problems in the latest changes.
EOF

opencode run "Add input validation to src/api.ts, then ask @reviewer to check it"
opencode run --agent plan "Outline how to split src/api.ts into modules"
opencode serve --port 4096     # headless server for @opencode-ai/sdk

• Remove OpenCode versions older than 0.1.x before installing (README).
• The TUI needs a modern terminal emulator; the docs list recommended ones.
• Most permissions default to allow (only doom_loop and external_directory default to ask, and .env reads are denied); set permission rules before letting it run unattended.
• Since v1.1.1 the old tools boolean config is deprecated in favour of permission; it still works for backwards compatibility.
• Every enabled MCP server adds its tool definitions to the context; the docs warn that some, such as the GitHub MCP server, can exceed the context limit.

Yes, as a client: local and remote MCP servers are added under mcp in opencode.json, and opencode mcp auth handles OAuth. A2A and AG-UI support are not documented.

The agent is MIT-licensed and works with your own provider keys. The same team sells optional model services (OpenCode Zen pay-as-you-go, OpenCode Go at $10 or $40 a month) and an Enterprise offering priced on request.

Primary agents such as Build and Plan handle the conversation and can call subagents (General, Explore, Scout or your own) through the Task tool; each call runs in a child session you can open.

OpenCode is MIT-licensed and works with many model providers; it also exposes an HTTP server and TypeScript SDK and runs as an ACP agent in editors.

TypeScript; its SDK is a JS/TS package, @opencode-ai/sdk.