OpenCode hit 199,000 GitHub stars as of this week - more than any other coding agent, open or closed. The Pinggy guide that tracks this space calls it "the open-source breakout of 2026." That is a real number, and it deserves a real answer: is it earned?
The short version is that the star count measures something genuine but not what most coverage implies. Provider neutrality and an LSP feedback loop are the actual reasons to reach for it. The benchmark gaps are also real. Both facts fit in the same tool.
What "open source coding agent" actually means in 2026
A coding agent is not just a model in a terminal. The surrounding framework - often called the harness - manages everything the model cannot do alone: reading files, executing shell commands, editing repositories, running tests, and feeding results back so the model can continue working step by step. That harness is what differs between tools, and it is where OpenCode makes its bets.
OpenCode moves up the field not just because of growth, but because of what it isn't: caught up in the acquisition wave that swallowed Cursor and Windsurf, or shut down like Gemini CLI. It crossed 165k GitHub stars (now 199k per the repo's release history) and the team rebranded to Anomaly, so the canonical repo is now anomalyco/opencode.
If you have old Docker image references pointing at sst/opencode, they will break.
The appeal is direct: model-agnostic, terminal-native, fully open source, and you own your data.
It supports 75+ LLM endpoints out of the box: Anthropic, OpenAI, Google, AWS Bedrock, Azure OpenAI, OpenRouter, and anything OpenAI-compatible including local Ollama servers.
The LSP feedback loop is the real technical differentiator
OpenCode spawns Language Server Protocol servers and feeds compiler diagnostics back to the model after every edit. If the agent introduces a type error, the next round includes that error, and the model self-corrects.
Agentic coding tools like Claude Code can write, refactor, and debug across an entire codebase, but by default they read code as plain text, the way grep does. The Language Server Protocol changes that: it's the same code-intelligence layer an IDE uses, and wiring it into an agent lets it read code by meaning instead of by string match.
OpenCode ships integration with 40-plus LSP servers - Pyright for Python, rust-analyzer, gopls, clangd, the TypeScript server, and more. The agent gets real diagnostics, hover info, go-to-definition, find-references, and call hierarchies straight from the same servers your editor uses. When OpenCode edits a function signature, it can see the compiler errors it just created and fix them in the same loop.
There is a catch worth knowing: the current LSP page says it is disabled by default. You have to wire it up yourself. If you install OpenCode and run it without touching the config, you are not getting this feature.
Sessions persist locally in SQLite, so a terminal session can be closed and resumed later with full conversation history and agent state intact. Code intelligence comes from LSP servers already installed in the project - the agent reads diagnostics and type information the same way an IDE would, instead of relying purely on the model's own reasoning about unseen code.
Beagle in action#eng-infra, 3:47pmThe 78% speed gap is real, and here is what causes it
A Builder.io head-to-head test in early 2026 using Claude Sonnet 4.5 on identical tasks found Claude Code consistently faster on the same model: OpenCode took nearly twice as long overall, but generated 29% more tests.
The extra time came from OpenCode running full test suites and safety checks by default. Whether that trade-off is worth it depends on your tolerance for regressions versus speed.
The architectural reason: OpenCode runs roughly 78% slower per task than Claude Code when both are using Anthropic's models. The gap comes from client-server overhead in OpenCode's architecture, since the desktop app communicates with a backend server even when running locally.
Here is what independent benchmark data shows on a repository modernization task:
| Agent | Model | Pass rate | Avg time (min) |
|---|---|---|---|
| Claude Code | Opus 4.5 | 48.2% | 38.5 |
| OpenCode | Opus 4.5 | 42.0% | 21.6 |
| OpenCode | GPT-5.2 | 43.0% | 16.4 |
| Codex CLI | GPT-5.2 | 30.4% | 20.3 |
Claude Code achieves the highest pass rate at 48.2%, followed by OpenCode with GPT-5.2 at 43.0%, OpenCode with Claude at 42.0%, and Codex CLI at 30.4%. Build success rates are high across all agents (≥95%), indicating that agents reliably produce compilable code even when functional correctness varies.
The non-obvious read on that table: OpenCode with GPT-5.2 finishes faster and scores higher than OpenCode with Claude on this benchmark. Provider flexibility is not just a philosophical stance - it has measurable task-level consequences.
The ecosystem around it: Goose at the Linux Foundation, Gemini CLI gone
OpenCode is not the only harness worth tracking, but the field narrowed fast. In December 2025, Block contributed Goose as a founding project of the Linux Foundation's Agentic AI Foundation (AAIF), alongside Anthropic's Model Context Protocol and OpenAI's AGENTS.md. The project's repository and governance were transferred from Block to the AAIF in April 2026.
Goose matters less for what it does than for how it is owned: it is the first major coding-adjacent agent to move from a corporate parent to a foundation, an exit path no other open harness offers. That is a meaningful governance signal for teams running regulated environments. Pricing is fully free and open-source under Apache 2.0 with no paid tier. Goose follows a BYOK model: users pay only for the LLM API usage they choose.
Meanwhile, Aider went quiet, with no commits since May 22, 2026. It remains the best Git-native editing model in the field, but teams should plan around the risk.
Roo Code is archived. The VS Code extension shipped its final release on May 15, 2026 and the repo is read-only. The team pivoted to a cloud product called Roomote - do not start new work on it; migrate to Kilo Code or Cline.
The practical comparison for teams choosing an open harness today:
| Tool | License | Stars | Local model support | LSP | Notable limitation |
|---|---|---|---|---|---|
| OpenCode | MIT | ~199k | Yes (Ollama) | Yes (disabled by default) | 78% slower than Claude Code |
| Goose | Apache 2.0 | ~51k | Yes (Ollama) | No | No autocomplete; newer tool |
| OpenHands | MIT | Active | Yes (via LiteLLM) | No | Heavier setup |
| Aider | Apache 2.0 | Stale | Yes | No | No commits since May 2026 |
| Cline | Apache 2.0 | Active | Yes | No | VS Code only |
The honest read is that OpenCode's star count is a proxy vote for "I do not want to be locked into one vendor." It earns that vote. Whether it earns your production workflow depends on whether the 78% speed penalty and the manual LSP setup are prices you will actually pay.
Open source coding agent: common questions
What is OpenCode and how does it differ from Claude Code?
OpenCode is a MIT-licensed, model-agnostic coding agent that runs in the terminal, desktop app, or IDE extension. Unlike Claude Code, which only works with Anthropic models, OpenCode supports 75+ LLM providers including local Ollama inference. Claude Code posts higher benchmark scores on most published evals; OpenCode's advantage is flexibility and data ownership.
Is OpenCode actually faster or slower than Claude Code?
Slower. On identical tasks with the same model, OpenCode runs roughly 78% slower than Claude Code. The gap comes from client-server architectural overhead and OpenCode's default behavior of running full test suites. It does generate more tests per run - whether that trade-off is worth it depends on your tolerance for CI failures.
Does OpenCode support local models and offline use?
Yes. OpenCode runs local inference through Ollama, llama.cpp, and any OpenAI-compatible endpoint. Your code and conversation history stay in a local SQLite database. If you point it at a cloud provider, that traffic leaves your machine - but the harness itself does not require cloud connectivity.
What is the Goose coding agent and who maintains it now?
Goose is an Apache 2.0 agent runtime originally built by Block and transferred to the Agentic AI Foundation at the Linux Foundation in April 2026. It runs entirely on-machine, supports 25+ model providers, and costs nothing beyond your API tokens. It is the only major open coding harness governed by a foundation rather than a company.
Does OpenCode's LSP integration actually work out of the box?
No. LSP is disabled by default in current builds. You need to configure it manually, pointing OpenCode at LSP servers already installed for your project (Pyright, rust-analyzer, gopls, etc.). When wired up, it gives the agent compiler-accurate diagnostics mid-loop - OpenCode's most underrated technical advantage on large refactors.