Use CoralBricks with OpenCode

Run OpenCode's coding agent on CoralBricks open-model inference — GLM and Kimi with up to 1M context. Set one environment variable and CoralBricks shows up in the model picker.

OpenCode is an open-source terminal coding agent. CoralBricks ships in OpenCode's built-in provider registry, so there is no config file to write: export your API key and the CoralBricks models appear in the picker. Chat, agent mode, and tool calling all route through the Coral gateway.

OpenCode listing the CoralBricks models straight from its provider registry, then running an agent task on GLM 5.2 — file written, tests run, all asserts passing

No config file: the models show up once CORAL_API_KEY is set, then opencode run does the work on GLM 5.2 — file writes, shell commands, streaming.

1. Get an API key

Create a key from your API keys page — it looks like cb_....

2. Connect

Pick either route; both end up in the same place.

Connect once (desktop app or CLI). In the desktop app, add CoralBricks from the provider list and paste the key. From the terminal:

opencode auth login

Either one saves the key to ~/.local/share/opencode/auth.json, which the desktop app and the CLI both read — connect in one and the other is connected too.

Or use the environment, which suits servers, CI, and throwaway shells:

export CORAL_API_KEY="cb_..."

Add that to your shell profile (~/.zshrc / ~/.bashrc) so it persists. Note that a GUI app launched from the Dock does not inherit your shell profile, so the desktop app needs the connect-once route (or a launch from a terminal that already has the variable).

3. Run it

opencode                                             # interactive TUI
opencode run "explain this repo" -m coralbricks/glm-5.2-fp4
opencode run "fix the failing test" -m coralbricks/kimi-k3

Switch models any time inside the TUI with /models, or set a default in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "coralbricks/glm-5.2-fp4"
}

opencode models | grep coralbricks lists the three models, but it lists them from the public registry whether or not you are authenticated — so it confirms OpenCode can see CoralBricks, not that your key works. The first opencode run is what proves the connection.

Available models

Model	Slug	Context	Input $/M	Output $/M
GLM 5.2	`coralbricks/glm-5.2-fp4`	1M	$1.12	$4.40
Kimi K3	`coralbricks/kimi-k3`	1M	$3.00	$15.00
GPT-OSS 120B	`coralbricks/gpt-oss-120b`	128K	$0.12	$0.60

Cached input tokens are always free.

Already added CoralBricks by hand?

Earlier versions of this guide had you define the provider yourself in ~/.config/opencode/opencode.json. That block is no longer needed, and if you leave it in you will see CoralBricks listed twice in the model picker — once from your config, once from the registry. Delete the provider entry (and point model at coralbricks/glm-5.2-fp4) to clear the duplicate:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "coralbricks/glm-5.2-fp4"
}

Notes

Tool calling and agent mode work out of the box — the gateway emits OpenAI-exact streaming tool calls; OpenCode's file edits, shell commands, and multi-step agent loops run unmodified.
Reasoning control: models think by default. Disable per request with "reasoning": {"enabled": false} if you call the API directly; OpenCode manages this itself.
Long context is the point — GLM 5.2 and Kimi K3 take up to 1M tokens, so whole-repo context and long agent sessions fit without pruning.
The API is OpenAI wire-format at https://inference.coralbricks.ai/v1 — the same config shape works in any tool that speaks OpenAI-compatible providers (Cline, Continue, aider). See the API reference for the full surface.