OpenCode
by Anomaly (Anomaly Innovations)
The open-source AI coding agent that runs anywhere and talks to any model
OpenCode is a free, MIT-licensed AI coding agent that runs in the terminal, a desktop app, a web UI or an IDE and works with roughly 75 model providers instead of locking you to one vendor. It is aimed at engineering teams that want an agentic coding assistant they can read, self-host and point at their own models or gateway.
OpenCode is an open-source AI coding agent built by Anomaly, the team behind the SST framework, and released publicly on 19 June 2025. It runs as a terminal UI, a CLI, desktop applications for macOS, Windows and Linux, a web interface and IDE extensions, with GitHub and GitLab integrations, and it is deliberately provider-agnostic: it reaches around 75 model providers through the models.dev catalogue rather than binding to a single vendor, so a team can point it at frontier APIs, at a corporate gateway, or at a locally hosted open-weight model. The agent ships two default modes, a build agent with full write access and a read-only plan agent for analysis, plus configurable sub-agents, per-tool permission rules, MCP and LSP server integration, code-formatter config, AGENTS.md project instructions, shareable session links, and an SDK, plugin API and standalone server component for embedding it in internal automation. It is MIT-licensed with roughly 202,000 GitHub stars, 26,200 forks, 950 contributors and more than 15,500 commits, making it one of the most-starred open-source coding agents in existence, and it installs via curl, npm, Homebrew, Scoop, Chocolatey or Pacman. Monetisation sits beside the free client rather than inside it: OpenCode Zen is a pay-as-you-go gateway of 60-plus models the team has benchmarked for coding, metered per million tokens with auto-reload and per-workspace spend caps. In March 2026, following legal requests from Anthropic, maintainer Dax Raad merged PR #18186 removing the Claude Pro/Max OAuth plugin, the Claude system prompt and the Claude Code beta header; Claude models remain reachable through ordinary API keys and through Zen.
Platform and developer-experience leads who need an agentic coding assistant they can audit, self-host and route through their own model gateway instead of a single vendor's cloud.
A production-grade coding agent with no per-seat licence — cost is limited to the model tokens you actually route through it.
At a Glance
- Category
- Developer Tools
- Pricing
- Free, Usage-based
- Target Market
- CTOs, Enterprise Developers, Platform Engineers, DevOps Engineers
- Deployment
- Open-source, Self-hosted, API-based
- Founded
- 2025
Key Features
- ✓Provider-agnostic model routing
Reaches roughly 75 LLM providers through the models.dev catalogue, so you can switch frontier, gateway or local models without changing tools.
- ✓Separate build and plan agents
A full-access build agent and a read-only plan agent mean exploratory analysis cannot mutate the repository by accident.
- ✓Permission rules and AGENTS.md
Per-tool permission configuration and committed project instructions constrain what the agent may run and read in each repository.
- ✓MCP and LSP integration
Connects to MCP servers for external tools and auto-configures language servers so the model sees real type information.
- ✓Multi-surface clients
One agent across terminal UI, CLI, macOS/Windows/Linux desktop apps, a web UI, IDE extensions and GitHub/GitLab integrations.
- ✓OpenCode Zen model gateway
A curated pay-as-you-go gateway of 60-plus coding-benchmarked models with per-workspace and per-member monthly spend caps.
- ✓SDK, plugins and server mode
An SDK, plugin API and standalone server let teams script the agent or embed it inside internal developer platforms.
Capabilities
Use Cases
- •Multi-model bake-offs
Run the same refactor through several frontier models from one client to compare output quality and token cost.
- •Self-hosted coding assistance
Point the agent at a locally served open-weight model so source code never leaves the network perimeter.
- •Large refactors and migrations
Use plan mode to review a change strategy on a read-only pass before the build agent touches any files.
- •Token-budgeted team rollout
Give each engineer a Zen workspace with a hard monthly spend cap instead of buying fixed per-seat licences.
- •Repository and CI automation
Drive the CLI and server component from GitHub or GitLab pipelines to triage issues and open pull requests.
Ideal For
Best For
- ✓Terminal-first engineers who want an agentic coding assistant without leaving the shell
- ✓Teams standardising on an internal LLM gateway that need a client accepting any provider endpoint
- ✓Privacy-sensitive or disconnected work where a locally hosted model must serve the agent
- ✓Cost-controlled AI coding rollouts billed per token rather than per developer seat
- ✓Comparing several frontier models on the same task without changing tooling
Not Ideal For
- ✗Organisations that need a vendor support contract and an SLA — this is a community project with a Discord, no published enterprise support tier, and no SOC 2 or ISO attestation
- ✗Teams hoping to fund agent usage from an existing Claude Pro or Max subscription; Anthropic's March 2026 legal requests removed that authentication path outright
- ✗Security-conservative environments that cannot absorb a fast-moving agent with ~4,100 open issues and a documented history of an unauthenticated local RCE and a bypassable command-permission model
- ✗Non-technical users who want a GUI-first assistant with no configuration files
Integrations
Deployment
Market Analysis
Pros
- ✓MIT-licensed with no per-seat cost, so spend tracks actual token consumption rather than headcount
- ✓Genuinely provider-agnostic across roughly 75 providers including local models, which removes vendor lock-in
- ✓One of the largest open-source agent communities — about 202,000 GitHub stars, 26,200 forks and 950 contributors
- ✓A read-only plan agent plus per-tool permission rules give reviewable control over what the agent is allowed to do
Cons
- ✗A January 2026 disclosure documented an unauthenticated local HTTP server exposing shell execution, interactive terminals and arbitrary file reads; before v1.0.216 any website could reach it through the browser, and the maintainers shipped silent fixes without a security advisory despite disclosure attempts starting November 2025
- ✗A July 2026 practitioner teardown found the bash permission system bypassable via environment variables, aliases, base64 encoding and subprocess calls, with grants persisting across sessions
- ✗Anthropic's March 2026 legal requests forced removal of Claude Pro/Max OAuth (PR #18186 drew 548 thumbs-down), so existing Claude subscriptions can no longer fund usage
- ✗The same teardown criticises the documentation as written for agents rather than humans, and notes the default model URL is fetched from models.dev so the tool phones out even on first launch
- ✗Roughly 4,100 open issues, no SLA, no SOC 2 or ISO 27001 attestation and no documented enterprise support offering
Pricing
OpenCode (open source)
$0
- ✓MIT-licensed client
- ✓Terminal, desktop, web and IDE surfaces
- ✓Bring your own provider API keys
- ✓~75 supported model providers
- ✓MCP, LSP, SDK and plugin support
OpenCode Zen (pay-as-you-go)
From $0.20 per 1M input tokens
- ✓60+ coding-benchmarked models
- ✓Per-workspace and per-member monthly spend caps
- ✓Auto-reload billing below a $5 balance
- ✓Rotating free trial models
The client is free and MIT-licensed, so the only cost is model tokens — paid either straight to a provider on your own API key or through OpenCode Zen, which meters per million tokens from $0.20 in / $1.20 out for a budget model, $2.00 / $10.00 for Claude Sonnet 5, $5.00 / $25.00 for Claude Opus 5, up to $30.00 / $180.00 for GPT 5.5 Pro. Zen auto-charges $20 when a balance drops below $5, passes card fees through at cost (4.4% + $0.30), and supports workspace and per-member monthly limits. There is no per-seat licence and no published enterprise price.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 5 on domains other than opencode.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Momentic
Agentic QA platform that writes, runs and self-heals end-to-end tests for web and mobile apps
Raindrop
AI agent monitoring that catches silent production failures: Sentry for AI agents
GitLab Duo Agent Platform
Agentic AI across the whole GitLab DevSecOps lifecycle: planning, coding, code review, CI/CD and security agents under one governance model
CodeRabbit
AI code review and agentic change management for teams shipping human- and machine-written code