Apertis MCP Server

@apertis/mcp-server is an MCP server that lets agents (Claude Code, OpenCode, or any MCP-compatible client) manage your Apertis account and query the API gateway at runtime — list models, check quota, create API keys, ask for task-aware model recommendations, and delegate bulk grunt work to a cheap coworker model.

GitHub: apertis-ai/apertis-mcp · npm: @apertis/mcp-server

Installation

No install step required — the server runs via npx. Just point your MCP client at it.

Claude Code

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "apertis": {
      "command": "npx",
      "args": ["@apertis/mcp-server"],
      "env": {
        "APERTIS_API_KEY": "sk-your-key-here"
      }
    }
  }
}

OpenCode

Add to openclaw.json:

{
  "mcpServers": {
    "apertis": {
      "command": "npx",
      "args": ["@apertis/mcp-server"],
      "env": {
        "APERTIS_API_KEY": "sk-your-key-here"
      }
    }
  }
}

Restart your agent and the Apertis tools become available.

Environment Variables

Variable	Required	Default	Description
APERTIS_API_KEY	Yes	—	Your Apertis API key (sk-...)
APERTIS_BASE_URL	No	https://api.apertis.ai	API base URL
APERTIS_COWORKER_MODEL	No	deepseek-v4-flash	Intern model used by the delegate tool

Available Tools

Tool	Description
list_models	List available models with optional filtering (free/paid, capability)
get_model_info	Get detailed info about a specific model (pricing, context, provider)
compare_models	Side-by-side comparison of 2–5 models
check_quota	Check account balance, subscription status, remaining quota
get_usage_stats	Usage statistics by model and period (today/week/month)
list_api_keys	List API keys with status and quota (masked)
create_api_key	Create a new API key with optional quota limit
suggest_model	Freeform keyword-based model search over the full catalog
recommend_model	Curated Apertis pick for a task type with live pricing
delegate	Offload bulk grunt work (file reads, boilerplate, summarization) to a cheap coworker model — keeps your Claude usage limit intact

recommend_model — Dynamic Model Selection

Prefer recommend_model when the task fits one of the 5 canonical types — it wraps the GET /v1/recommend endpoint and returns a single opinionated pick rather than a ranked list.

Input

Parameter	Required	Description
task	Yes	coding / long-context / fast-chat / reasoning / vision
budget	No	low / medium (default) / high

Example conversation

You:   What model should I use for coding right now?
Agent: [calls recommend_model(task="coding", budget="medium")]
       → Apertis recommends claude-sonnet-4-6
         Input: $2.40/1M, Output: $12.00/1M
         Why: Best coding ability per dollar. 200K context.
         Alternatives:
         - deepseek-v3 ($0.30/1M) — 3x cheaper, good for simpler coding
         - claude-opus-4-6 ($4.00/1M) — most capable, higher cost

Use the returned model value directly in subsequent API calls.

delegate — Offload Grunt Work to a Cheap Coworker

delegate gives your agent a cheap "intern". Claude stays the manager on your Anthropic subscription and hands off high-volume, low-judgement work — bulk file reads, boilerplate generation, summarization — to a cheap model through the Apertis gateway. The coworker reads files itself, so their bulk content never enters your agent's context, and the work runs on Apertis credit instead of your Claude weekly/5-hour limit.

Input

Parameter	Required	Description
instruction	Yes	What the intern should do
file_paths	No	Local file paths the coworker reads itself; content never enters your context
content	No	Inline content to process
model	No	Override the intern model (default deepseek-v4-flash)

The coworker calls are ordinary Apertis requests billed against your API key — one credit pool, transparent per-call usage in your dashboard. Unreadable file paths and gateway errors are surfaced as tool errors, never silently swallowed.

Example conversation

You:   Summarize the errors in build.log
Agent: [calls delegate(instruction="Summarize the errors in <=10 bullets",
        file_paths=["build.log"])]
       → returns a concise summary; the full log never enters Claude's context

Auto-delegation. To make your agent delegate without being asked each time, paste the routing rules from coworker-rules.md into your project's CLAUDE.md. They tell the agent when to delegate (bulk reads, boilerplate, summarization) versus do the work itself.

Resources

In addition to tools, the server exposes three read-only resources:

Resource URI	Description
apertis://account	Current account info, balance, subscription
apertis://models	Full model catalog (cached 5 min)
apertis://usage/today	Today's usage summary

Security

• All API keys in tool output are masked (first 4 + last 4 characters)
• New keys are shown in full once — save them immediately
• Keep APERTIS_API_KEY out of source control; use your MCP client's env-var expansion