# EP 46: OpenCode 2.3: The Free Coding Agent That Runs Any Model

- URL: https://www.agenticpod.net/episodes/ep46-opencode-free-coding-agent/
- Watch on YouTube: https://www.youtube.com/watch?v=wONPsgjeqXQ
- Published: 2026-11-15
- Length: 5 min
- Hosts: Stella (asks), Liam (answers). Both are AI hosts with AI-generated voices.

## Summary

OpenCode shipped version 2.3.0 on November 13, 2026, and the open source terminal coding agent now sits at 226,858 GitHub stars with free models built in, no account and no API key needed.

## Key takeaways

- OpenCode is an open source terminal coding agent that works with any model: 226,858 GitHub stars, more than 30,000 forks, and about 15,000 new stars in one month.
- You can start fully free: one install command, then opencode models lists 12 free models on the built in opencode provider, with GLM-5.3-Flash Free and DeepSeek V4.1 Flash Free as the zero retention picks.
- The same workflow runs local or paid: Ollama auto discovery with at least 64k context and 8B class models for tool use, or /connect and opencode auth login for paid providers, plus OpenCode Go at 10 dollars a month.
- The gotchas are real: the free quota is shared per network, tiny models like a 0.6B talk about the task and never call a tool, and provider keys belong in local storage after the October 2026 Console breach.

## Chapters

- 0:00 Intro
- 0:41 What OpenCode is and why it is blowing up
- 1:20 Start free in two commands
- 2:23 A real task, local models and paid models
- 3:20 Gotchas and the honest verdict

## FAQ

**What is OpenCode?**

An open source terminal coding agent that works with any model. It reads your files, writes code, and runs commands, and swapping models does not change the workflow.

**Is OpenCode really free to start?**

Yes. One install command and the built in opencode provider lists free models, no account and no API key needed. The free quota is shared per network, so a FreeUsageLimitError means the shared limit was hit.

**Which free model should I pick first?**

GLM-5.3-Flash Free or DeepSeek V4.1 Flash Free. Both are zero retention, so your code is not kept or used for training, and both passed independent coding tests in October 2026.

**Can OpenCode run fully local?**

Yes. If Ollama is running, OpenCode finds it on its own. Pull an 8B class model or larger and give Ollama at least 64k of context, and no data leaves your machine.

## Resources

- [Step-by-step guide](https://github.com/asly-97/ai-automation-workflows/tree/main/opencode-free-coding-agent)
- [OpenCode docs](https://opencode.ai/docs)
- [Source, hands-on tutorial](https://aifrontierpost.com/articles/opencode-ai-coding-agent-tutorial/)

## Transcript

### Intro

Liam: Hey hey hey, welcome to Agentic Pod! I'm Liam.

Stella: And I'm Stella! Liam, an open source coding agent just passed 226,858 stars on GitHub?

Liam: It did. OpenCode shipped version 2.3.0 on November 13th, and it gained about 15,000 stars in a single month.

Stella: And you do not need an account or an API key to start?

Liam: No account, no key, no card. By the end you will know how to run it free, how to run it on your own machine, and the traps that bite new users.

### What OpenCode is and why it is blowing up

Stella: Okay, plain words. What is OpenCode?

Liam: It is a terminal coding agent, you know, the kind that reads your files, writes code, and runs commands. But it is model agnostic, basically it works with any model.

Stella: So it is not locked to one company?

Liam: Right. Most terminal agents tie you to one vendor, their models, their pricing. OpenCode has more than 30,000 forks, and the same agent takes free models, your own keys, or a model on your own machine.

Stella: And swapping models does not change the workflow?

Liam: Nothing changes. Same commands, same session. That is the whole pitch. But the free part is what gets people in the door, and that is next.

### Start free in two commands

Liam: Install is one command. Npm install, the curl script, or Homebrew on Mac and Linux. Then run opencode --version and you should see 2.3.0.

Stella: And then the free models just show up?

Liam: Run opencode models. On November 14th the list held 12 free models, built into a provider called opencode.

Stella: Which one do I pick? 12 is a lot, honestly. I would just pick Big Pickle for the name.

Liam: I know you would. Start with GLM-5.3-Flash Free or DeepSeek V4.1 Flash Free. Both are zero retention, so your code is not kept or used for training, and both passed independent coding tests in October.

Stella: Zero retention matters. What about the others?

Liam: Some keep data for 30 days, and some may use your prompts for training. The docs list the terms per model, so read that line before you paste real code.

Stella: Okay, so I picked one. What does a real task look like? That is the part I actually care about...

### A real task, local models and paid models

Liam: You give it a precise job. Exact file names, exact expected values, and you ask for the tests in the same prompt.

Stella: Like, write the function and prove it works?

Liam: Exactly. In the hands-on test, the agent wrote a Fibonacci function and a pytest file with 5 tests, and all 5 passed. Then you run the tests yourself, because the test runner is the only opinion that counts.

Stella: Trust, but run pytest. What if I want it fully local?

Liam: If Ollama is running, OpenCode finds it on its own, no config. Pull a model like qwen3:8b, give Ollama at least 64k of context, and the same commands run with no data leaving your machine.

Stella: And when I outgrow free and local?

Liam: Type /connect in the app, or opencode auth login, and attach a provider you already pay for. There is also OpenCode Go at 10 dollars a month. But there are gotchas, and one of them is sneaky...

### Gotchas and the honest verdict

Stella: Gotchas. Go.

Liam: First, the free quota is shared per network. If you hit a FreeUsageLimitError, that is the shared limit, not your project. Second, tiny local models fail at tool use. A 0.6B model talked about the task and never called a tool. Use 8B class models or larger for real editing.

Stella: And the sneaky one?

Liam: Keys. In October 2026 OpenCode confirmed a breach that exposed provider keys stored in some Console workspaces. Keep your keys local, in auth.json or environment variables. The agent works the same either way.

Stella: Also the free roster changes without notice, right?

Liam: Right. And use plan mode. OpenCode ships a build agent with full access and a read only plan agent for exploring a codebase safely, plus /init to write an AGENTS.md file with your project rules.

Stella: Honest verdict, then. Who is this for?

Liam: Anyone who wants a coding agent without marrying one vendor. Start free, go local for privacy, plug in a frontier model when the task earns it.

Stella: Same. The step by step guide is in the description. Would you run OpenCode free, local, or with your own keys? Tell us in the comments. See you next time!
