Skip to main content

OpenCode

OpenCode is an open-source AI coding agent for the terminal. Nebul's Inference API is OpenAI-compatible, so you can register it as a custom provider in OpenCode and use any model your project has access to.

Prerequisites​

Installation​

Install OpenCode using one of the methods below. For Windows, Docker, and other options, see the OpenCode install guide.

bash
1
curl -fsSL https://opencode.ai/install | bash

Configure Nebul​

OpenCode stores provider credentials separately from provider configuration. Follow both steps below.

1. Add your API key​

  1. Start OpenCode in your project directory:
bash
1
opencode
  1. Run /connect, scroll to Other, and enter a provider ID — for example nebul.
  2. Paste your Nebul API key when prompted.

The provider ID must match the key you use in opencode.json in the next step. Credentials are stored in ~/.local/share/opencode/auth.json.

2. Configure the provider​

Create or update opencode.json in your project root (or ~/.config/opencode/opencode.json for a global default). Add Nebul as an OpenAI-compatible provider:

json
12345678910111213141516171819202122
{
  "$schema": "https://opencode.ai/config.json",
  "model": "nebul/zai-org/GLM-5.1-FP8",
  "provider": {
    "nebul": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Nebul Inference API",
      "options": {
        "baseURL": "https://api.inference.nebul.io/v1"
      },
      "models": {
        "zai-org/GLM-5.1-FP8": {
          "name": "GLM 5.1 FP8",
          "limit": {
            "context": 128000,
            "output": 16384
          }
        }
      }
    }
  }
}

Replace the model key under models with any model ID from your Model Catalog. Set limit.context and limit.output to match the model's actual capabilities when possible — OpenCode uses these values for context management.

If you used a different provider ID in /connect, replace nebul everywhere in this config with that same ID.

3. Select a model​

Inside OpenCode, run:

1
/models

Your Nebul provider and configured models appear in the list. Select the model you want to use, then start coding.

For first-time project setup, run /init to generate an AGENTS.md file that helps OpenCode understand your codebase. See the OpenCode providers documentation for advanced options such as custom headers and environment-variable API keys.

Troubleshooting​

  • Provider not listed — Confirm the provider ID in /connect matches the key under provider in opencode.json.
  • Model not found — The model key in config must exactly match the ID Nebul expects. Check the Model Catalog.
  • Connection errors — Verify the base URL is https://api.inference.nebul.io/v1 and that your API key is valid for the active AI Studio project.
info

Config changes may require restarting OpenCode. If a model still misbehaves after updating limits, restart the TUI and run /models again.

Tips​

Optional Nebul-specific notes. Everything else — MCPs, plugins, multi-agent setups, config precedence — is covered in the OpenCode docs.

Context & sessions​

Long sessions fill up fast. Make sure auto-compaction is enabled:

bash
1
unset OPENCODE_DISABLE_AUTOCOMPACT

Add that to your shell profile to persist it. Use /new and /session to manage separate conversation contexts. Switch models mid-session with /model.

Permissions​

OpenCode allows most tool use by default. For daily work, prompt before destructive actions — at minimum set "bash", "edit", "write", and "webfetch" to "ask" in your permission block. See OpenCode config for the full list of keys.

json
12345678910111213141516
"permission": {
  read: allow,
  glob: allow,
  grep: allow,
  list: allow,
  bash: ask,
  edit: ask,
  write: ask,
  task: ask,
  skill: ask,
  webfetch: ask,
  websearch: ask,
  codesearch: ask,
  external_directory: ask,
  doom_loop: ask
}

Model stack​

OpenCode works well with a primary + lightweight model pair. A common Nebul setup:

ModelRole
zai-org/GLM-5.1-FP8Primary — planning, refactoring, tool use
MiniMaxAI/MiniMax-M2.5Fast tasks — set as "small_model"

Add other models from your Model Catalog as needed. Store keys in env vars (NEBUL_API_KEY) and reference them in config with {env:NEBUL_API_KEY} — see custom providers.

Multi-agent workflows (optional)​

Plugins like micode and oh-my-opencode split work into roles (planner, implementer, reviewer) and can route each role to a different Nebul model. Install via ~/.config/opencode/package.json, map agents in micode.json, and add the plugin name to "plugin" in opencode.json.

Web search via Nebul​

For docs lookup and research, Firecrawl MCP is a good default. Point it at Nebul's endpoint:

json
123456789
"firecrawl": {
  "type": "local",
  "command": ["npx", "-y", "firecrawl-mcp"],
  "enabled": true,
  "environment": {
    "FIRECRAWL_API_KEY": "{env:FIRECRAWL_API_KEY}",
    "FIRECRAWL_API_URL": "https://firecrawl.inference.nebul.io"
  }
}

Use Playwright MCP only when you need full browser interaction (logins, SPAs). See OpenCode MCP configuration for setup details.

Project instructions​

Run /init to generate a starter AGENTS.md. OpenCode also reads CLAUDE.md automatically. Use these files for stack constraints, test commands, and coding conventions per repo.

Further reading​

Last updated on