OpenCode - Open Source AI Coding Agent CLI

OpenCode sits in the same agentic CLI neighborhood as Codex CLI, Pi, NanoClaw, and Herdr, but its shape is different.

It is not just a terminal command. The current anomalyco/opencode repository is a Bun/TypeScript monorepo with a core CLI, server package, web app, desktop app, plugin package, SDK, LLM/provider layer, terminal UI pieces, and packaging infrastructure.

OpenCode is an open-source AI coding agent that can run as a terminal coding assistant, a local server, a web interface, and a desktop app.

OpenCode GitHub Source Code OpenCode Documentation OpenCode Desktop Download License: MIT

What is OpenCode?

OpenCode is a coding-agent CLI for local developer workflows.

The quick install options from the README include:

curl -fsSL https://opencode.ai/install | bash
npm i -g opencode-ai@latest
brew install anomalyco/tap/opencode

Once installed, the main command is:

opencode

The command surface I validated locally includes:

opencode --help
opencode run "explain this repository"
opencode serve --hostname 127.0.0.1 --port 14096
opencode web
opencode models openai
opencode providers
opencode agent list
opencode mcp
opencode session

The installed binary on this machine was opencode 1.15.13. The cloned source checkout was newer, at 1.16.2.

That difference matters: I validated the installed CLI and server behavior, but I did not validate a fresh source build of the newest checkout.

Repository Shape

The repository I inspected was:

tmp/foss-post/opencode

Source snapshot:

Important paths:

This is also an important SEO and reader-clarity note: some older search results and articles refer to different or stale “OpenCode” projects. The current repo covered here is the Bun/TypeScript anomalyco/opencode project with the opencode-ai npm install path.

Built-in Agent Modes

The README documents two built-in agents:

It also documents a general subagent that can be invoked as:

@general

That separation is useful because many coding-agent workflows benefit from a read-only planning mode before handing execution to a higher-permission agent.

Provider and Model Support

OpenCode is provider-flexible. The provider/model spec in this checkout includes provider IDs such as:

• opencode
• anthropic
• openai
• google
• google-vertex
• github-copilot
• amazon-bedrock
• azure
• openrouter
• mistral
• gitlab

I also validated the installed CLI model listing for OpenAI:

opencode models openai

It returned many OpenAI model IDs, including gpt-5, gpt-5-codex, gpt-5.3-codex, gpt-5.4, gpt-5.5, image models, embedding models, and older GPT-4/o-series entries.

Provider support is not the same thing as being ready to spend tokens immediately. You still need credentials and provider configuration.

Local Field Test

Local environment during the trial:

Source Dev Path

The cloned repo expects Bun:

bun install
bun dev

But on this machine:

bun --version

failed with:

bun: command not found

So I did not run the repo-local bun install, bun dev, or package tests from source.

That is the main local blocker for contributors: install Bun 1.3+ before trying to develop directly from the monorepo.

Installed CLI Path

This machine already had OpenCode installed:

which opencode
opencode --version
opencode --help

Observed:

/home/jalcocert/.nvm/versions/node/v22.22.0/bin/opencode
1.15.13

The help output exposed commands for completion, ACP, MCP, TUI mode, run, providers/auth, agents, server mode, web mode, model listing, stats, session export/import, plugins, database tools, and GitHub workflows.

Local Server UI

The documented server default is port 4096.

My first attempt failed:

opencode serve --hostname 127.0.0.1 --port 4096 --print-logs

because port 4096 was already in use by another local process:

Failed to start server. Is port 4096 in use?

I retried on a high port:

opencode serve --hostname 127.0.0.1 --port 14096 --print-logs

Result:

opencode server listening on http://127.0.0.1:14096

Fetching the local URL returned the OpenCode web app shell, including the root HTML, JavaScript bundle, stylesheet, favicon references, and a #root application mount point.

One important warning appeared:

Warning: OPENCODE_SERVER_PASSWORD is not set; server is unsecured.

For anything beyond a local-only quick test, set a server password and avoid binding it publicly without understanding the auth model.

Docker Notes

There is a packages/opencode/Dockerfile, but this repo is not a compose-first self-hosted web app in the same way as a dashboard or media service.

The Dockerfile expects a prebuilt OpenCode binary under a dist/ path and packages that binary into an Alpine image. I did not validate a Docker Compose deployment, so I did not add a Home-Lab compose file for this post.

For readers, the practical path today is:

• install the released CLI
• run the TUI locally
• run opencode serve or opencode web if you want the local web surface
• use source checkout only after installing Bun

Development Commands

From the repo docs:

bun install
bun dev
bun dev <directory>
bun dev serve
bun dev serve --port 8080
bun run --cwd packages/app dev
bun run --cwd packages/desktop dev

Tests should not be run from the root package:

bun test

The root package intentionally guards against that with:

do not run tests from root

Run tests and typechecks from package directories instead, for example:

bun run --cwd packages/opencode test
bun run --cwd packages/opencode typecheck

Where It Fits

I would put OpenCode in this mental model:

OpenCode is closest to Codex CLI and Pi, but with its own provider system, local app/server pieces, and desktop/web ambitions.

The honest pick:

• Pick OpenCode if you want an open-source coding agent with a local server, web surface, desktop ambitions, and provider flexibility.
• Pick Codex CLI if you want the OpenAI-native path and the reference Codex developer workflow.
• Pick Pi if you want a reusable TypeScript agent toolkit as much as a CLI.
• Add Herdr if your problem is supervising several terminal agents, not choosing one agent.

FAQ