OpenCode

Anatomy of the Fastest-Growing Open Source AI Coding Agent

Deep Dive

github.com/anomalyco/opencode

The Numbers

Zero to 100K stars in 8 months

102K

GitHub Stars

732

Contributors

9,130

Commits

912

Releases

1M

Lines of Code

10M+

Total Downloads

First commit: March 21, 2025 — MIT License — TypeScript + Bun

The Origin Story

From Serverless Stack to AI Coding Agents

2017 Jay V & Frank Wang meet at University of Waterloo, launch Anomaly Innovations

2021 Take SST (Serverless Stack) through Y Combinator — $1M raised from PayPal, LinkedIn, YouTube founders

2025 Q1 SST turns profitable with 25K GitHub stars. Dax Raad (early SST user, joined 2021) becomes co-founder

Jun 19, 2025 OpenCode launches — Jay, Frank, Dax & Adam Elmore ship the open-source AI coding agent

Jan 2026 Hits 100K stars, 650K MAU, adopted by Cloudflare — 300K+ downloads/day

The Thesis

"There will be an explosion of AI coding models. Developers will demand open-source alternatives to locked-in tools."

— Jay V, CEO of Anomaly

The Gap They Saw

Claude Code locked to Anthropic. Cursor locked to its IDE. No open-source terminal agent that could work with any model, any provider.

The Explosion

Total Downloads Over Time

58K

Jun

261K

Jul

475K

Aug

764K

Sep

1.19M

Oct

1.78M

Nov

2.77M

Dec

10.2M

Jan 26

Jan 6, 2026: 247K downloads in a single day. By Jan 29: 386K/day. The hockey stick went vertical.

System Architecture

Client-Server with Provider-Agnostic Core

ClientsTUI (SolidJS) / Desktop (Tauri) / Web (Astro) / IDE Extensions
ServerHono HTTP + WebSocket API on port 4096
Agent EngineVercel AI SDK — build, plan, general agents
Providers + Tools75+ models / 46+ tools / LSP + MCP protocols
Storage + InfraFile-based sessions / SST on Cloudflare + PlanetScale

Key insight: the server is the brain. Clients are thin — TUI, desktop, and web all talk to the same API.

Monorepo: 18 Packages

One repo, many surfaces

Package	Role
opencode	Core CLI + server + agent
app	Shared web UI (SolidJS)
desktop	Native app (Tauri v2)
web	Landing page (Astro 5)
ui	Component library
sdk	Auto-generated API client
plugin	Public extension API

Package	Role
console/*	User accounts + billing
function	Webhooks + GitHub App
identity	Auth (OAuth2)
slack	Slack bot integration
enterprise	Enterprise features
extensions	VS Code / IDE plugins
containers	Docker configs

Build orchestration: Turbo — Infrastructure: SST — Runtime: Bun 1.3+

The Core Engine

Agent Loop: Stream → Tool → Permission → Repeat

// Simplified agent loop
while (response.hasToolCalls) {
  for (const call of response.toolCalls) {
    // Check permissions
    const ok = await permission.ask(call)
    if (!ok) throw PermissionDenied

    // Execute tool
    const result = await tool.run(call)
    history.push(result)
  }
  // Stream next LLM response
  response = await provider.stream(history)
}

Three Agent Types

build — Full access, default agent for development
plan — Read-only, analysis and exploration
general — Sub-agent for complex multi-step tasks

Switch agents with Tab. Each has its own permission ruleset, model, and temperature.

Provider System

75+ models, 20+ providers, zero lock-in

Bundled Providers

Anthropic (Claude) OpenAI (GPT-5+) Google (Gemini) Azure OpenAI Mistral Groq Amazon Bedrock DeepInfra XAI (Grok) Cohere OpenRouter GitHub Copilot TogetherAI Cerebras GitLab AI Local models

OpenCode Zen

Curated provider with hand-tested deployments per model. Volume discounts flow back to users.

Powered by Vercel AI SDK — unified interface for all providers with streaming, tool use, and caching.

Use a cheap model for planning, then switch to an expensive model for execution.

Tool System: 46+ Tools

Zod-validated, permission-gated, LSP-integrated

Category	Tools
File Ops	read, write, edit, patch, glob, grep, ls
Execution	bash, pty, task (sub-agents)
Search	websearch, codesearch, sourcegraph
Analysis	lsp, diagnostics
Custom	User-defined .ts/.js in project
External	MCP servers (stdio + SSE)

Safety Layers

Read-before-write enforcement
Modification time verification
Uniqueness checks for edits
Diff preview before approval
LSP diagnostics after every write

Tool.define("edit", {
  parameters: z.object({
    path: z.string(),
    old_string: z.string(),
    new_string: z.string()
  }),
  execute(args, ctx) { /* ... */ }
})

TUI: Built for Terminal Power Users

SolidJS + OpentUI = reactive terminal UI

SolidJS signals for reactive state (not React)
OpentUI — their own terminal UI library
Context providers: routing, SDK, theme, keybindings, sync
11 dialog types: permissions, sessions, command palette, model picker...
Prompt history with frecency ranking

Desktop wraps the same UI via Tauri v2 with 12 native plugins (clipboard, notifications, auto-update, etc.)

// TUI entry point (simplified)
const App = () => (
  <RouteProvider>
    <SDKProvider>
      <ThemeProvider>
        <KeybindProvider>
          <DialogProvider>
            <SessionView />
          </DialogProvider>
        </KeybindProvider>
      </ThemeProvider>
    </SDKProvider>
  </RouteProvider>
)

Development Velocity

Commits per month — acceleration is the story

152

Apr

301

May

654

Jun

650

Jul

614

Aug

518

Sep

709

Oct

1,049

Nov

1,877

Dec

2,083

Jan

912 releases in 10 months = ~3 releases/day

Peak: 176 releases in June alone

Most active: Tuesday (1,748 commits)

Peak hour: 12:00 PM (lunchtime shipping)

AI-Powered Development

They eat their own dogfood

Co-Authored-By Breakdown

GitHub Action

981 commits

rekram1-node

217

opencode-agent

100

OpenCode

82

Claude

19

AGENTS.md

They maintain a dedicated style guide for AI agents working on the codebase:

Single-word variable names
No try/catch, use .catch()
No else, use early returns
No destructuring — use dot notation
Prefer const + ternaries over let
Use Bun.file() APIs

Code Review & Community

High trust, low ceremony, strict quality gates

PR Patterns (last 100 merged)

Merged without formal review (84%)

Had code reviews (16%)

Core team ships fast with high trust. Community PRs get reviews.

Trust & Vouch System

vouch @user — mark as trusted
denounce @user — auto-close issues/PRs
No AI-generated walls of text in PRs
Issue-first policy (no PRs without issues)
Conventional commits required

Denouncement is reserved for repeated low-quality AI spam, not disagreements.

Permission & Security Model

Granular control over every agent action

// Permission ruleset
{
  "read": { "*": "allow",
             "*.env": "ask" },
  "write": { "*": "ask" },
  "bash": { "*": "allow",
             "rm -rf": "deny" },
  "external_directory": "ask",
  "doom_loop": "ask"
}

Three modes: per-action, session-wide, auto-approve (CI/CD)

Privacy-First Architecture

No code stored — zero telemetry by default
Sessions stored locally as JSONL
.env files always require explicit approval
External directories gated separately
Doom loop detection prevents infinite agents

Cloudflare adopted OpenCode for internal use precisely because of this architecture.

Session Management

Compaction, forking, sharing, undo

Session Lifecycle

Create — new conversation, optional parent for threading
Compact — summarize old messages, shrink context window
Fork — branch a session for exploration
Share — encrypted URL for read-only replay
Revert — undo to any message with file snapshots

Message Format

type MessageV2 =
  | Human    // user input + attachments
  | Assistant // multi-part response
    { content: Part[]
      // Text, ToolCall, Error, Meta
      thoughts?: string
      moderation?: Status }
  | System   // system prompts

Stored as JSON at ~/.opencode/message/{sid}/{mid}.json

OpenCode vs Claude Code

Daniel Miessler: "Claude Code's secret sauce isn't so secret"

	OpenCode	Claude Code
Models	75+ models, 20+ providers	Claude only
Lock-in	None — MIT licensed	Anthropic ecosystem
UI Surfaces	TUI + Desktop + Web + IDE	CLI only
Architecture	Client-server (remote-able)	Monolithic CLI
Privacy	Zero telemetry, local storage	Anthropic cloud
Strength	Thoroughness, provider flexibility	Polish, orchestration quality

Both tools now deliver comparable results on complex development workflows. The market is expanding beyond a single dominant solution.

Dax Raad's Philosophy

What the creator thinks you're getting wrong about AI

"Feeling productive and being productive are fundamentally different things."

Sequential beats parallel — one task with a fast model often beats multi-agent chaos
Codebase quality > model quality — clean types and file structure improve LLM performance more than model upgrades
LSP is the superpower — immediate compiler feedback lets agents self-correct
Benchmarks are broken — "If you can only notice your LLM is better on a benchmark, the end user can't tell"

Lessons for Teams

How to ship like the OpenCode team

Velocity Tactics

Ship daily — 912 releases in 10 months
Trust core team — 84% PRs merge without review
Automate everything — 981 GitHub Action commits
Dogfood aggressively — use your own agent to build your agent
AGENTS.md — give AI contributors explicit style rules

Quality Gates

Issue-first — no PR without a linked issue
Conventional commits — structured history
Denounce spam — Vouch system blocks bad actors
No AI walls of text — respect reviewer time
Feature gates — UI/core changes need design review

The formula: High trust for the core team + strict gates for the community + AI agents as first-class contributors.

The Agent is
Open Source

OpenCode shipped the future of coding agents — in the open, at breakneck speed.

agent37.club

github.com/anomalyco/opencode • opencode.ai

OpenCode

The Numbers

Zero to 100K stars in 8 months

The Origin Story

From Serverless Stack to AI Coding Agents

The Thesis

The Gap They Saw

The Explosion

Total Downloads Over Time

System Architecture

Client-Server with Provider-Agnostic Core

Monorepo: 18 Packages

One repo, many surfaces

The Core Engine

Agent Loop: Stream → Tool → Permission → Repeat

Three Agent Types

Provider System

75+ models, 20+ providers, zero lock-in

Bundled Providers

OpenCode Zen

Tool System: 46+ Tools

Zod-validated, permission-gated, LSP-integrated

Safety Layers

TUI: Built for Terminal Power Users

SolidJS + OpentUI = reactive terminal UI

Development Velocity

Commits per month — acceleration is the story

AI-Powered Development

They eat their own dogfood

Co-Authored-By Breakdown

AGENTS.md

Code Review & Community

High trust, low ceremony, strict quality gates

PR Patterns (last 100 merged)

Trust & Vouch System

Permission & Security Model

Granular control over every agent action

Privacy-First Architecture

Session Management

Compaction, forking, sharing, undo

Session Lifecycle

Message Format

OpenCode vs Claude Code

Daniel Miessler: "Claude Code's secret sauce isn't so secret"

Dax Raad's Philosophy

What the creator thinks you're getting wrong about AI

Lessons for Teams

How to ship like the OpenCode team

Velocity Tactics

Quality Gates

The Agent isOpen Source

The Agent is
Open Source