OpenCode MCP server

OpenCode Primer

Everything you need to know about OpenCode — from your first prompt to custom agents, skills, plugins, and MCP integrations. Clear mental models, real examples, every fact verified against the current release.

curl -fsSL https://opencode.ai/install | bash

Who this is for: Developers using (or about to use) OpenCode. Beginners get a guided path; power users get depth on Custom Commands, Skills, Plugins, MCP, and Agents.

⚖️ Not affiliated with the OpenCode team. This is a community-maintained guide. For canonical sources, check opencode.ai/docs and github.com/anomalyco/opencode.

💡 Pro Tip: If this guide saves you an afternoon, a ⭐ helps other developers find it.

---

🧭 Choose your path

You are…	Start with	Time
🚀 New to OpenCode	What is it → Setup → Prompt Engineering	~15 min
⚡ Already using it	Custom Commands · Skills · Plugins · MCP	~20 min each
🧠 Configuring or headless	Agents · Headless & CI · Models & Providers	varies

---

Getting started

What is OpenCode?

OpenCode is an open-source AI coding agent — a terminal app (TUI), desktop app, and IDE extension that reads your repo, runs commands, edits files, and talks to any LLM you point it at. Maintained by Anomaly, MIT-licensed — and as of July 2026 the most-starred coding agent on GitHub (182K+ stars), with 8M monthly users by the founder's count.

Three things it does that a chat UI can't:

• Reads your actual repo — not pasted snippets. Greps your files, follows imports, grounds answers in real context using built-in read/grep/glob tools.
• Edits in place and runs your stack — diff-aware edit/write/apply_patch, then bash to run your tests, linter, or build on the spot.
• Composes with the rest of your toolchain — multi-provider models, custom slash commands, agent skills, JS/TS plugins, MCP servers, LSP, formatters, and an HTTP server for CI use.

Three interfaces:

Surface	Command	When
TUI (default)	opencode	Interactive day-to-day work in your terminal
CLI / headless	opencode run "<prompt>"	Scripts, CI jobs, one-shot prompts
Server	opencode serve or opencode web	Headless API, web UI, or remote attach

opencode                                       # start TUI in the current repo
opencode run "fix the failing test in src/api.test.ts"
opencode serve --port 4096                     # headless server

OpenCode sits in the same space as Claude Code, Codex CLI, and Cursor — same problem, different trade-offs. None is universally better; pick the one whose model, surface, and ecosystem fit your workflow. (The field consolidated hard in 2026: Gemini CLI was retired in June in favor of the closed-source Antigravity CLI, and Roo Code shut down in May.)

⚠️ AI-coding caveat: OpenCode (like every coding agent) can produce wrong code, miss edge cases, hallucinate APIs, and over-apply patterns. You're still the reviewer. Read diffs before accepting, run tests, and don't auto-approve destructive operations on code you care about.

---

When OpenCode isn't the right tool

Limitation	Detail
No inline editor completion	OpenCode is a conversational agent, not Copilot-style autocomplete. For ghost-text-while-you-type, reach for Copilot, Cursor Tab, Windsurf (formerly Codeium), or Supermaven — or alongside.
Quality depends on the underlying model	OpenCode doesn't replace the LLM's reasoning. If your task fails on Sonnet, switching to OpenCode won't fix it.
Provider-agnostic ≠ provider-equivalent	Some models tool-call better than others. A slug swap isn't free.
TUI on slow SSH / minimal terminals	The TUI uses truecolor and complex layouts. Degrades over slow connections. opencode serve + attach or opencode run are better for those cases.
Plugins run arbitrary code	Any plugin — local or npm — has full user permissions. Audit before installing.
Free Zen models are time-limited	The Zen docs call them "available for a limited time." Don't build production on them.
Docs lag shipping	OpenCode moves fast. This guide and even the official docs occasionally trail the actual binary.
Not a substitute for code review	Diff-aware edits + passing tests don't guarantee correct, secure, or maintainable code. Treat AI output as a junior teammate's PR.

When something else fits better:

• Want a polished single-vendor product → Claude Code or Codex CLI.
• Deeply embedded in VS Code workflow → Cursor or GitHub Copilot.
• Want a small Python tool you can read end-to-end → Aider (note: development has slowed markedly in 2026).
• Don't want to manage provider keys → a managed hosted product, or OpenCode Go.

OpenCode's trade-off: flexibility and openness over polish and single-vendor integration. Worth it for many use cases; not all.

---

Setup

⏱️ 5 minutes from zero to first AI-assisted commit.

1. Install

curl -fsSL https://opencode.ai/install | bash

Other supported installers (pick whichever you already use):

npm i -g opencode-ai@latest         # npm / pnpm / yarn / bun all work
brew install anomalyco/tap/opencode # macOS / Linux
scoop install opencode              # Windows
choco install opencode              # Windows
sudo pacman -S opencode             # Arch
paru -S opencode-bin                # Arch (AUR)
mise use -g github:anomalyco/opencode      # mise users
docker run -it --rm ghcr.io/anomalyco/opencode  # Docker

A desktop app for macOS / Windows / Linux is in beta at opencode.ai/download.

2. Authenticate

opencode auth login                  # interactive provider picker
opencode auth list                   # see who you're signed in to

Or from inside the TUI: /connect. Provider options covered in the next section.

3. First prompt

cd ~/your-project
opencode

Try one of these:

• explain what this codebase does
• add a README section about installation
• find and fix the failing test in src/api.test.ts

4. Generate an AGENTS.md

/init

/init generates an AGENTS.md — your project's "house rules" that OpenCode reads every session. Commit it. More in Prompt Engineering.

📝 Coming from Claude Code? OpenCode reads CLAUDE.md as a fallback if no AGENTS.md exists. Disable with OPENCODE_DISABLE_CLAUDE_CODE=1 once migrated.

5. (Bonus) See a real configured project

This repo's own .opencode/ directory is a working example:

Path	What it does
.opencode/agents/	Custom agent definitions (markdown + YAML)
.opencode/commands/	Custom slash commands
.opencode/plugins/	Working JS plugins (audit log, secret blocker)
.opencode/skills/	Agent skills with SKILL.md files
AGENTS.md	Project instructions OpenCode reads every session
opencode.json	Project config with safe defaults

---

Models & Providers

OpenCode is provider-agnostic. Four ways to plug in a model — pick whichever fits your accounts, budget, and privacy posture:

	Best for	Trade-off
Direct providers (Anthropic, OpenAI, Google, Groq, OpenRouter, AWS Bedrock, Azure)	You already have credits with a provider	Manage one key per provider
OpenCode Zen — pay-as-you-go gateway	One key, curated lineup, no commitment	Pricier per token than going direct in some cases
OpenCode Go — flat-rate subscription	Predictable $10/month, zero key management	Curated open models only, usage limits
Local models (Ollama, LM Studio, llama.cpp)	Offline, strict data-residency, hobbyist	Quality varies hugely with model and hardware

There's no universally right answer — pick what fits.

⚠️ Warning: You can no longer sign in with a Claude Pro/Max subscription. OpenCode removed its built-in Anthropic OAuth login in March 2026 at Anthropic's legal request (PR #18186). Anthropic models still work fine — via an API key in the provider config below, or via Zen.

Direct provider config

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anthropic/claude-sonnet-5",
  "provider": {
    "anthropic": { "options": { "apiKey": "{env:ANTHROPIC_API_KEY}" } },
    "openai":    { "options": { "apiKey": "{file:~/.secrets/openai-key}" } }
  }
}

Variable substitution: {env:NAME} (environment) or {file:./path} (file content, absolute or relative).

Per-agent model override:

{
  "agent": {
    "plan":         { "model": "anthropic/claude-haiku-4-5" },
    "deep-thinker": { "model": "openai/gpt-5.5", "options": { "reasoningEffort": "high" } }
  }
}

OpenCode Zen

A curated, pay-as-you-go AI gateway. The OpenCode team tests and tunes models for coding-agent workloads; you get one key and a benchmarked lineup. Reference Zen models with the opencode/<model-id> namespace.

# 1. Sign up at https://opencode.ai/auth, add billing, copy API key
# 2. In OpenCode:
/connect            # pick "OpenCode Zen", paste key
/models             # browse the lineup

Pricing snapshot (verified 2026-07-05; check opencode.ai/docs/zen for current)

Model ID	Input / Output (per MTok)	Reach for it when…
opencode/claude-fable-5	$10 / $50	Absolute frontier — the hardest reasoning problems
opencode/claude-opus-4-8	$5 / $25	Flagship reasoning — complex refactors, large multi-file work
opencode/claude-sonnet-5	$2 / $10	Balanced everyday coding — newer and cheaper than Sonnet 4.6
opencode/claude-haiku-4-5	$1 / $5	Fast, lightweight tasks
opencode/gpt-5.5 (≤272K)	$5 / $30	Long-context, OpenAI ecosystem
opencode/gpt-5.4-mini	$0.75 / $4.50	Cost-efficient frontier
opencode/gemini-3.1-pro (≤200K)	$2 / $12	Multimodal, Google ecosystem
opencode/qwen3.7-plus	$0.40 / $1.60	Budget-friendly heavy lifting

Free tier (rotating, time-limited, for community feedback): big-pickle (stealth), deepseek-v4-flash-free, mimo-v2.5-free, north-mini-code-free, nemotron-3-ultra-free.

📚 Full lineup, cached-read/write rates, GPT-5 Codex variants, Gemini, GLM, Kimi, MiniMax, DeepSeek, Grok, and the Claude Opus 4.8/4.7/4.6/4.5 line: opencode.ai/docs/zen · Deeper guide (incl. deprecation sunset dates): docs/zen.md.

💡 Pattern: cheap explorer / strong executor. Use a cheap model (qwen3.7-plus, gpt-5.4-mini) for explore and scout subagents that read a lot, and a stronger one (claude-sonnet-5 or claude-opus-4-8) for the main build agent that does the editing.

OpenCode Go

The flat-rate alternative to Zen: $10/month ($5 your first month) for a curated set of hosted open coding models — no API keys, no per-token math. Usage limits apply (roughly $12 per 5 hours / $30 per week / $60 per month of equivalent usage); past them, Go can fall back to a Zen balance.

Zen vs. Go in one line: Zen bills per token across the full lineup; Go is a fixed subscription to open models. Reach for Go to learn and hack on a predictable budget; reach for Zen (or direct keys) when you want frontier models. Details: opencode.ai/docs/go · docs/zen.md.

---

Using OpenCode

Prompt Engineering

📖 Project Initialization. Run /init to auto-generate an AGENTS.md. Treat it like any frequently-used prompt — iterate on it. Don't just dump 500 lines and forget.

Plan → Build cycle

Two primary agents you toggle with Tab:

Agent	What changes	Use it for
plan	edit and bash default to ask — analysis without surprise writes	Reading code, designing approaches, reviewing diffs
build	Full tool access (default)	Implementation, refactors, running tests

Natural workflow:

[plan]  > how would you implement OAuth for this app?
       (reads, plans, proposes — no writes)
[plan]  > <Tab>
[build] > do it
       (implements, edits, runs tests)

💡 Prompts like think hard, think more, or ultrathink nudge reasoning depth on supported models — the same pattern in the Anthropic prompt engineering guide.

File references and shell

Syntax	Effect	Example
@<path>	Attach file content (fuzzy resolved)	look at @src/api/auth.ts
!<cmd>	Run shell, inject output	!git diff then ask about it

> review @packages/api/src/handlers/auth.ts for missing input validation
> !pnpm test --reporter=verbose
> based on those failures, fix the auth handler

AGENTS.md for project rules

# AGENTS.md

## Commands
- build: `pnpm build`
- test:  `pnpm test`
- lint:  `pnpm lint && pnpm typecheck`

## Conventions
- TypeScript strict mode, no `any`
- Prefer named exports; default exports only for React components

## Don't
- Edit anything under `vendor/`
- Run migrations without confirmation
- Push to `main` directly

Reads on every session. Add extra files via instructions: ["..."] in opencode.json.

---

TUI & Themes

Leader key

OpenCode's TUI uses a leader key (default ctrl+x) followed by a single letter. Keeps single-keystroke editing keys free for the input box.

Binding	Action
ctrl+x n	New session
ctrl+x l	List sessions
ctrl+x c	Compact session
ctrl+x e	Open external editor
ctrl+x m	Browse models
ctrl+x t	Switch theme
ctrl+x u / ctrl+x r	Undo / Redo
ctrl+x q	Quit
ctrl+p	Command palette
Tab	Cycle primary agents
Esc	Cancel / dismiss

Readline shortcuts also work inside the prompt box (ctrl+a, ctrl+e, ctrl+k, ctrl+u).

Themes

Built-in: system · tokyonight · everforest · ayu · catppuccin · catppuccin-macchiato · gruvbox · kanagawa · nord · matrix · one-dark. Set via /themes or tui.json.

{
  "$schema": "https://opencode.ai/tui.json",
  "theme": "tokyonight",
  "leader_timeout": 2000,
  "keybinds": { "command_list": "ctrl+p" }
}

Custom themes: .opencode/themes/<name>.json (project) or ~/.config/opencode/themes/<name>.json (global). Truecolor terminals required.

📚 Deep dive: docs/tui.md · opencode.ai/docs/keybinds · opencode.ai/docs/themes.

---

Headless & CI

Three ways to run OpenCode unattended:

# 1. One-shot prompts
opencode run "summarize the last 5 commits"
opencode run --format json "list every TODO with file and line" > todos.json
opencode run --agent plan --model anthropic/claude-haiku-4-5 \
  "audit src/ for missing input validation"

# 2. Headless server / web UI
opencode serve --port 4096 --hostname 0.0.0.0
opencode web   --port 4096

# 3. GitHub agent
opencode github install     # add a workflow file to your repo
opencode pr 123             # check out PR #123 and start a session

Useful run flags: --continue / --session <id> (resume), --fork (branch a session), --share (publish), --format json (machine-readable), --file <path> (attach), --replay (interactive replay, v1.16+), --auto (auto-approve everything not deny — CI only). There's also a lightweight --mini mode (v1.17.10+).

Sessions, stats, sharing:

opencode session list                  # past sessions
opencode stats --days 30 --models      # token & cost by model
opencode export <id> --sanitize        # JSON dump (with redaction)
opencode import session.json           # restore from JSON or share URL

📚 Full CLI reference: docs/reference/cli.md. CI recipes (GitHub Actions, GitLab CI): docs/workflows.md.

---

Sharing

/share generates a public link to the current session (opncd.ai/s/<id>). Three modes via opencode.json:

{ "share": "manual" }     // default — only when you ask
{ "share": "auto" }       // every new session is shared
{ "share": "disabled" }   // forbid sharing entirely

/unshare revokes the link. Enterprise builds can enforce "disabled" via managed config.

⚠️ Privacy: Shared conversations live on OpenCode's servers until you unshare. Don't share sessions with secrets, proprietary code, or confidential data.

---

Extending OpenCode

OpenCode has five extension points. Pick by trigger and surface:

flowchart TD
    Start([Repeatable task you want<br/>OpenCode to do]) --> Q1{How does it trigger?}
    Q1 -->|User types /name| Q2{Just a prompt template,<br/>or a full role<br/>with permissions?}
    Q1 -->|Model auto-discovers<br/>from a description| AS[Agent Skill<br/>.opencode/skills/NAME/SKILL.md]
    Q1 -->|Event in OpenCode<br/>itself| Q3{Local code or<br/>external system?}
    Q2 -->|Prompt template| CC[Custom Command<br/>.opencode/commands/NAME.md]
    Q2 -->|Full role| CA[Custom Agent<br/>.opencode/agents/NAME.md]
    Q3 -->|Local JS/TS| PL[Plugin<br/>.opencode/plugins/NAME.js]
    Q3 -->|External tool| MCP[MCP Server<br/>opencode.json → mcp]

Extension	Triggered by	Lives in
Custom Commands	User typing /<name>	.opencode/commands/*.md
Agent Skills	Model auto-discovery (description match)	.opencode/skills/<name>/SKILL.md
Plugins	Lifecycle events (tool calls, file edits, sessions)	.opencode/plugins/*.{js,ts}
Agents	Tab key or @<name> mention	.opencode/agents/<name>.md
MCP Servers	Prompt-driven external tool use	opencode.json → mcp

💡 These five compose. Most polished workflows combine 2–3.

---

Slash Commands

OpenCode ships a set of built-in slash commands plus everything you write yourself.

Command	What it does
/init	Guided generation of AGENTS.md
/help	Show the help dialog
/new (/clear)	New session
/sessions (/resume, /continue)	List and switch sessions
/models	Browse models
/connect	Add a provider
/themes	Switch theme
/compact (/summarize)	Compact session to save tokens
/share · /unshare	Public link control
/export	Export to markdown
/undo · /redo	Walk edit history
/editor	Open $EDITOR
/details · /thinking	Toggle visibility of tool / reasoning blocks
/exit (/quit, /q)	Exit

Custom commands can override built-ins with the same name.

📚 Full reference: docs/reference/slash-commands.md.

---

Custom Commands

Mental model: A custom command is a prompt template you've named. Save as .opencode/commands/<name>.md, invoke as /<name> in the TUI — with $ARGUMENTS, @file, and !shell interpolated at invocation time.

Minimal example:

---
description: Run tests and triage failures
agent: build
---

!`pnpm test --reporter=verbose`

If any tests failed above, open the failing files and propose precise fixes.
Otherwise, summarize the coverage.

Save as .opencode/commands/test.md → invoke with /test.

Frontmatter key	Purpose
description (required)	Shown in the slash-command palette
agent	Which agent runs it (build, plan, or custom)
model	Override the model just for this command
subtask	true runs as a subagent (isolated child session)

Body placeholder	Substitutes
$ARGUMENTS / $1 / $2	Text after the command name
@<path>	File content (fuzzy resolved)
!`<cmd>`	Shell stdout (backticks required in templates)

This repo ships four examples in .opencode/commands/: /review, /pr, /test, /optimize.

📚 Deep dive: docs/commands.md · opencode.ai/docs/commands.

---

Agent Skills

Mental model: A skill is a discoverable workflow. The agent sees the skill's name and description and decides on its own to load it when the task matches. You don't have to remember to invoke — the model does.

OpenCode uses Claude Code-compatible Agent Skill folders. Lookup paths (project → global → Claude-compat → Agent-compat):

.opencode/skills/<name>/SKILL.md
~/.config/opencode/skills/<name>/SKILL.md
.claude/skills/<name>/SKILL.md
.agents/skills/<name>/SKILL.md

Each SKILL.md needs YAML frontmatter with a name (lowercase-kebab, ≤64 chars) and a description (1–1024 chars). The description is the discovery signal — write it like a trigger sentence the agent will recognize.

Command vs. skill, in one line: commands are user-invoked (/<name>), skills are model-invoked (description match). Write both for the same workflow if you want both surfaces.

This repo ships one example: .opencode/skills/git-release/SKILL.md.

📚 Deep dive: docs/skills.md · opencode.ai/docs/skills.

---