OpenCode

Connect OpenCode to Mixen — pay from your balance in roubles, a terminal coding agent with a catalog of 105+ models.

Installation

npm install -g opencode-ai

Or with Homebrew (macOS/Linux):

brew install anomalyco/tap/opencode

Configuration

Create opencode.json in your project root (or globally at ~/.config/opencode/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "mixen": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Mixen AI",
      "options": {
        "baseURL": "https://api.mixen.ai/v1",
        "apiKey": "{env:MIXEN_API_KEY}"
      },
      "models": {
        "gpt-5.6-sol": { "name": "GPT-5.6 Sol", "id": "openai/gpt-5.6-sol" },
        "claude-opus-5": { "name": "Claude Opus 5", "id": "anthropic/claude-opus-5" },
        "glm-5.3": { "name": "GLM 5.3", "id": "z-ai/glm-5.3" },
        "glm-5.3-flash": {
          "name": "GLM 5.3 Flash",
          "id": "z-ai/glm-5.3-flash",
          "tool_call": true
        }
      }
    }
  },
  "model": "mixen/gpt-5.6-sol"
}

Key points:

• npm: "@ai-sdk/openai-compatible" — the driver for the /v1/chat/completions protocol that Mixen speaks.
• The key in models is a short name for OpenCode’s picker, while the id property is the identifier actually sent to the API. Mixen IDs contain a slash (openai/gpt-5.6-sol), so it goes into id, and the model field references the key: mixen/gpt-5.6-sol.
• A model’s entry in models accepts extra properties: "tool_call": true — tool-calling support (important for agent work), "reasoning": true — a reasoning-capable model; the allowed property names are listed in the config.json schema.
• {env:MIXEN_API_KEY} pulls the key from an environment variable, keeping it out of the git-tracked config. Alternatively run /connect → Other → ID mixen → paste the key: it is stored in a local auth file and picked up by matching the provider ID.

The API key environment variable (mxn-… keys are issued in the dashboard):

export MIXEN_API_KEY="your-api-key"   # add to ~/.bashrc or ~/.zshrc to persist

Running

opencode

The default model comes from the model field; switch it any time inside the session with the /models command.

Modes and agents

OpenCode has two work modes, toggled with the Tab key (the indicator sits in the lower-right corner):

• Build — the main mode: the agent edits files and runs commands;
• Plan — the agent only analyzes and proposes a change plan, without touching the project. Handy for running a complex task through Plan first, then repeating it in Build.

For routine jobs there are subagents — invoked by the primary agent or manually via @: @general for multi-step research, @explore for a fast read-only sweep of the codebase, @scout for read-only research of external docs and dependencies. Custom roles are added as files under .opencode/agents/ (project) or ~/.config/opencode/agents/ (global) — the filename becomes the agent name. Example .opencode/agents/review.md:

---
description: Review changes before a commit
model: mixen/glm-5.3-flash
mode: subagent
permission:
  edit: deny
  bash: ask
---

You are a strict reviewer. Check the changes for: leaked secrets,
broken error handling, non-obvious names. Answer with a list of
findings with file and line; if there are none, say so.

• model: mixen/glm-5.3-flash — each agent can have its own model; subagents inherit the calling agent’s model by default.
• permission allows or denies tools (edit, bash, webfetch and others; levels allow / ask / deny, with glob patterns like "git push": "ask").
• The file body below the frontmatter is the agent’s system prompt.

Useful session commands

Command	What it does
/models	Switch the model inside the session
/connect	Store a provider key in the local auth file
/init	Create an AGENTS.md in the project root — rules for the agent
/undo, /redo	Revert and re-apply the agent’s changes
/share	A link to the current conversation (conversations aren’t shared by default)

Typing @ opens a fuzzy search over the project’s files — and an image can be dragged into the terminal to add it to the context.

Cache and savings for agent sessions

An agent session is dozens of turns sharing the same growing prefix: the rules, the files read, the edit history. That prefix fits the prompt cache perfectly:

• repeated context within 5 minutes is billed at ~10% of the input price (for claude-sonnet-5 — 21.4 ₽ per 1M instead of 213.9 ₽, for glm-5.3-flash — 1.7 ₽), and every hit extends the window by another 5 minutes — a dense session stays cached from start to finish;
• a pause longer than 5 minutes resets the cache: the next turn pays the full input price, then the discount accumulates again;
• don’t edit AGENTS.md and rules mid-session — a prefix change resets the cache.

A typical role layout:

Role	Model	Price per 1M tokens
Primary agent: edits, hard tasks	anthropic/claude-sonnet-5	213.9 / 1069.5 ₽
Subagents: review, search, doc research	z-ai/glm-5.3-flash	8.4 ₽ input / 27.9 ₽ output

This way the subagent routine runs on a cheap model while a flagship makes the decisions. When calling the API directly, reasoning depth is regulated by reasoning_effort (from off to max; a model’s levels are in GET /v1/models → capabilities.reasoning_efforts), and remember that reasoning spends the answer budget.

Recommended models

Model	ID
GPT-5.6 Sol	openai/gpt-5.6-sol
Claude Opus 5	anthropic/claude-opus-5
GLM 5.3	z-ai/glm-5.3
DeepSeek V4 Pro	deepseek/deepseek-v4-pro
Kimi K3	moonshotai/kimi-k3

Full list — in the catalog. Add the models you need to the models section of the config as shown above.

Troubleshooting

• 401 / Invalid API Key — MIXEN_API_KEY is not set in the session where OpenCode runs (check with echo $MIXEN_API_KEY), or the key was copied incompletely.
• Model not found — the model field is built as mixen/<key from models>; the key is the short slash-free name, the slash stays inside id.
• Requests go to another provider — the selected model is not from mixen/*: switch it with /models or set the model field in the config.
• The agent doesn’t call tools — file and command support depends on tool calling in the specific model; try another one from the catalog and give it "tool_call": true in the config.
• Plan mode makes no edits — that’s expected: Plan only proposes changes; switch to Build with the Tab key to apply them.