A developer on a regulated fintech team opens a terminal, runs one command, and the agent reads the codebase, writes a multi-file diff, runs the compiler, reads the error back, and fixes its own mistake - without touching VS Code, without sending source code to a third-party server, and without a monthly seat fee beyond raw API tokens. That is OpenCode's pitch in one sentence, and it has resonated: the project crossed 160,000 GitHub stars in roughly 16 months.
Most coverage stops at the star count. This post digs into why the architecture works, where it genuinely breaks down, and what you should actually look at if you are evaluating open-source coding agents.
What OpenCode actually is (and what it is not)
OpenCode is a terminal-native coding agent, not an IDE extension. It runs in your shell and has no VS Code extension, no JetBrains plugin, no web UI.
It was built by the team behind SST (now Anomaly), runs locally on your machine, and connects to 75+ AI providers. Under the hood, it is a full agent harness - a tool loop, LSP diagnostics, and session management - not a chat wrapper.
The architectural choice that matters most is the Language Server Protocol integration. OpenCode spawns LSP servers and feeds compiler diagnostics back to the model after every edit. If the agent introduces a TypeScript type error, the next round of context includes the error and the model can self-correct. That is not the same as an agent that writes code and waits for a human to notice something is broken. The loop closes automatically.
It runs locally, stores conversations in SQLite, and gives you full control over which models process your code. The MIT license means you can inspect, modify, and self-host the entire system.
Why it grew past IDE-native agents
When OpenCode launched in March 2025 as a Go-based alternative to Aider and Cline, the conventional wisdom was that terminal-only was a deliberate niche - the kind of thing a few vim users would love and the rest of the market would ignore. The rest of the market did not ignore it.
Between January and April 2026, OpenCode crossed Cline, crossed OpenHands, and closed the gap on Aider despite Aider's two-year head start.
It hit 147,000 GitHub stars and 6.5 million monthly developers by April 2026, growing 4.5x faster than Claude Code in star velocity.
The reason is not UI flair. It is model agnosticism combined with privacy. According to the 2025 Stack Overflow Developer Survey, 84% of developers are using or planning to use AI tools in their workflow, yet 52% either avoid AI agents entirely or stick to simpler assistants - and a lot of that hesitation comes down to control over which model runs, cost when usage scales, and where your source code actually goes. OpenCode addresses all three in one binary.
One notable episode: Anthropic blocked OpenCode from accessing Claude models in January 2026. The issue was that OpenCode had been using the same OAuth authentication flow as Claude Code to access Claude Pro and Max subscriptions - Anthropic classified this as unauthorized spoofing of their official client and deployed server-side blocks on January 9.
The community quickly pivoted to GPT-5.4 and Gemini 3.1 Pro as recommended defaults. That a project could absorb losing access to its most popular model backend and keep growing is a real signal about ecosystem resilience.
Beagle in action#engineering, 10:47amThe honest tradeoffs
Setup friction is real: getting OpenCode configured - especially with local models - requires more technical knowledge than installing Cursor or Copilot. It is not beginner-friendly out of the box.
OpenCode does not provide inline IDE autocomplete like Copilot or Cursor. It operates at the task level, not the keystroke level. Many developers run both OpenCode and a Copilot or Cursor subscription simultaneously. That is not a failure mode - it is just a different contract with the user.
Here is a direct comparison worth having before you choose:
| Dimension | OpenCode | Claude Code | Cursor |
|---|---|---|---|
| License | MIT, fully open | Proprietary | Proprietary |
| Model support | 75+ providers | Anthropic only | Multi, IDE-native |
| Runs offline | Yes (Ollama) | No | No |
| IDE integration | Extension + terminal | Terminal | IDE-first |
| LSP self-correction | Yes | Yes | Yes |
| Code stored by vendor | No | Anthropic infra | Anysphere infra |
| Monthly seat cost | Token cost only | Subscription | Subscription |
Output quality depends entirely on the model you connect. A poorly-chosen or misconfigured model produces poor results. If you point OpenCode at a 7B quantized model to save money, you will get 7B-quality results. The agent harness does not compensate for a weak model.
What the benchmark numbers actually mean right now
Most coding agents advertise SWE-bench Verified scores. You should treat those numbers skeptically.
OpenAI has stopped reporting scores on SWE-bench Verified, the widely-used AI coding benchmark, after discovering that nearly 60% of problems its models failed contained fundamentally broken tests. OpenAI's February 23, 2026 analysis also found evidence that all major frontier models - including GPT-5.2, Claude Opus 4.5, and Gemini 3 Flash - had been trained on benchmark solutions, rendering scores meaningless.
SWE-bench Pro, the recommended replacement, paints a different picture: models scoring 80% on Verified dropped to roughly 23% on Pro - a benchmark designed to resist contamination.
SWE-bench Pro expands the benchmark to 1,865 long-horizon tasks across public, held-out, and commercial codebases, explicitly designed to reduce contamination and better reflect enterprise-level engineering work.
This matters directly for open-source agent evaluation. Anthropic's Opus 4.7 release notes describe screening SWE-bench Verified, Pro, and Multilingual for problems showing signs of memorization and excluding the flagged subset when reporting headline numbers. When vendors are actively curating which problems they report, you are not comparing apples to apples. OpenAI has flagged training data contamination concerns for SWE-bench Verified across all frontier models; SWE-bench Pro is emerging as the more reliable successor benchmark.
The honest signal for evaluating an agent like OpenCode is not a leaderboard score - it is what happens when you point it at your own codebase, with your preferred model, on a task that ships real code.
OpenCode open-source coding agent: common questions
What is OpenCode and how does it differ from Cursor?
OpenCode is a terminal-native, MIT-licensed coding agent that connects to 75+ LLM providers, including local models via Ollama. Cursor is an IDE fork with deep in-editor UX. OpenCode stores nothing on vendor servers and costs only API tokens; Cursor charges a monthly seat fee and requires their hosted infrastructure.
Does OpenCode work with local models?
Yes. OpenCode connects to Ollama and LM Studio out of the box, so it can run entirely offline with no data leaving your machine. Output quality scales with the model: a 34B model will outperform a 7B on complex refactors, but both run on your hardware.
Why did Anthropic block OpenCode in January 2026?
OpenCode had been using the same OAuth flow as Claude Code to access Claude Pro and Max subscriptions. Anthropic classified this as unauthorized spoofing of their official client, deployed server-side blocks on January 9, and a follow-up legal request in March 2026 resulted in OpenCode removing all Anthropic references from the codebase.
Is SWE-bench a reliable way to compare coding agents?
Not anymore on the Verified variant. SWE-bench Verified is increasingly contaminated, and OpenAI now recommends SWE-bench Pro. Models that scored ~80% on Verified score closer to 23% on Pro. Use benchmarks as a directional signal, not a hiring rubric.
What are OpenCode's real weaknesses?
Setup requires more technical knowledge than installing Cursor or Copilot, and there is no inline keystroke-level autocomplete. It is a task-level agent, not a completion engine. Teams that want both typically run OpenCode alongside a Copilot subscription for completions.