OpenCode kept showing up lately. Its launch post on Hacker News in March 2026 got over 1,200 points, it’s open source under the MIT license, and it has more than 210,000 GitHub stars (as of October 2026). Its main selling point is that it isn’t tied to any one vendor and lets you pick whatever model you want, so before actually installing it, I sat down and sorted out what it is.
What it is — compared with Claude Code
It’s a terminal-based AI coding agent built by Anomaly Innovations (the team behind the SST framework). It’s written in TypeScript and Bun, and its client-server design means the terminal, desktop app, and IDE extensions all attach to the same engine. It supports more than 75 inference providers (OpenAI, Anthropic, Google, DeepSeek, Groq, OpenRouter, even local Ollama), and has Plan/Build mode switching, AGENTS.md (the same role as Claude Code’s CLAUDE.md), and MCP server support.
Put it next to Claude Code and the directions split clearly. Claude Code uses only Anthropic models and is valued for how polished its agent runtime is; OpenCode trades that polish for freedom from lock-in.
Real-world criticism too — the prompt-sending controversy and the Grok rumor
I didn’t want to look only at the marketing copy, so I looked for real problems too.
- In v1.2.20, even with only local models configured, it was quietly sending prompts to
gpt-5-nanoon OpenCode’s own proxy (Zen) to generate session titles. A user caught this by watching the traffic with mitmproxy. It was a fallback to a free model when there was no small model for title generation, and v1.2.23 removed that fallback. (A claim that “it sends your prompts to Grok” also went around in the same discussion, but a maintainer denied it and it doesn’t match the measured record.) - For a terminal UI, issues about it using several GB of memory keep coming up.
- Some also pointed out that development moves so fast that features break between versions. That’s both an upside, since fixes land fast, and a downside, since things break often.
I checked three subscriptions, and only one worked — ChatGPT, Claude, and Gemini subscriptions
This was what I most wanted to know. If I already pay for an AI subscription, can I use it as-is in OpenCode, or do I have to buy a separate API key?
ChatGPT Plus/Pro works officially. Pick OpenAI in /connect and log in with ChatGPT Plus/Pro, and you can use GPT models right away on that subscription with no extra charge. OpenCode’s docs also list ChatGPT Plus as a subscription you can use with no setup.
Claude Pro/Max is different. You can connect it through community-made plugins, but even OpenCode’s own docs say “Anthropic explicitly prohibits this.” It’s an unofficial route with a real chance of violating policy.
Google AI Pro/Ultra used to work through an OAuth plugin, but since June 18, 2026, Google no longer serves requests from consumer plans (Gemini Code Assist for individuals, Google AI Pro, Google AI Ultra) in Gemini CLI at all. Only the business Gemini Code Assist subscriptions (Standard/Enterprise) remain, and even those need a Google Cloud project specified separately. Before that, Google had already named using Gemini CLI’s OAuth with third-party software as an example of a policy violation (a March 2026 announcement).
In short, of the three subscriptions only OpenAI can be used safely as-is; the other two mean accepting pay-as-you-go API key spending.
How it fits with Honcho and Hermes — adding memory to OpenCode over MCP
I checked what happens to Honcho, the AI memory service I use now (a service that keeps a record of conversations and pulls it up when needed). Honcho provides an official MCP server and OpenCode can attach any MCP server, so the connection itself works. The depth of integration is different, though. Unlike the way I use it now in Hermes, an always-on AI agent (set as the external memory provider in the dashboard so context gets inserted automatically), attaching it over MCP brings it in as “one tool the LLM calls when it decides it needs it.”
And Hermes Agent and OpenCode were never really comparable to begin with. Hermes is an always-on agent platform whose gateway is always running and which is managed from a dashboard (messaging gateway, scheduled skills, even multiple profiles); OpenCode is an interactive coding tool that only moves when you open a terminal. Whether it runs all the time is fundamentally different, so one can’t replace the other.
So I’m not switching for now
In the end, two things held me back.
One is memory. I keep building up my conversations in Honcho. Even when I’ve half forgotten what I did a while ago, I can search and find it. Switching to OpenCode doesn’t bring that accumulated context along, and the freedom to swap models doesn’t make up for that loss.
The other is remote control. If a tool only runs while I’m attached to a terminal, I still have to stay at my desk, exactly like with what I use now. As long as that constraint stays the same, there’s no reason to switch tools just for the choice of model.
In the end, the only criterion that decided whether to use this tool was “do I actually plan to mix and match several models?” Right now, I don’t.