Claude Code advanced

Claude Code is Anthropic’s terminal coding agent, and it talks to MirAPI over the Anthropic Messages API (POST /v1/messages). This page covers what to configure once the connection works: pinning the model, restricting permissions, wiring lifecycle hooks, and capping spend. There is no OpenAI or Gemini variant to set up — Claude Code only speaks the Anthropic protocol.

Prerequisites: Claude Code configured with ANTHROPIC_BASE_URL="https://api.mirapi.ai" (the bare origin, no /v1) and ANTHROPIC_AUTH_TOKEN — see Claude Code.

settings.json

Project settings live in .claude/settings.json in the project root. The keys you are most likely to touch:

Key	What it does
`model`	Pins the MirAPI model for the project, e.g. `deepseek-chat`
`max_turns`	Caps the tool-use turns per task, bounding runaway loops
`permissions.allow`	Tools and patterns Claude Code may run without asking
`permissions.deny`	Tools and patterns to block or flag
`hooks`	Commands that run on lifecycle events (`Stop`, `Notification`, `PostToolUse`, …)
`env`	Per-project environment variables (an alternative to shell exports)
`statusLine`	A custom command for the statusline

{
  "model": "deepseek-chat",
  "max_turns": 40,
  "permissions": {
    "allow": [
      "Read",
      "Edit",
      "Bash(git:*)"
    ],
    "deny": [
      "Bash(rm:*)",
      "Bash(curl:*)"
    ]
  },
  "hooks": {
    "Notification": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "terminal-notifier -message 'Claude Code finished' 2>/dev/null || true"
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "echo 'paused for input'"
          }
        ]
      }
    ]
  }
}

permissions is the whitelist of what Claude Code may do without confirmation; anything else triggers an interactive prompt.
hooks fire on lifecycle events and are the place for notifications, formatting, or guards.

Switching models

Change model in settings.json, or pass ANTHROPIC_MODEL at launch:

ANTHROPIC_MODEL="deepseek-chat" claude

The model name must match an entry from GET /v1/models exactly. A model outside the key’s whitelist returns 403. Claude Code is optimized for Anthropic-style clients, so pick a model that supports tool calling for the best experience.

Capping spend

The quota is enforced on the gateway side, so even a runaway session cannot exceed it:

In https://console.mirapi.ai → API Keys, set a quota limit (spend cap) on the key. Once the cap is reached, requests are rejected with 403 until you raise it.
Check usage at any time:

curl https://api.mirapi.ai/api/usage/token -H "Authorization: Bearer $MIRAPI_API_KEY"
curl https://api.mirapi.ai/api/log/token -H "Authorization: Bearer $MIRAPI_API_KEY"

GET /api/usage/token returns account-level totals; GET /api/log/token returns per-request detail.

Billing

Conversations are billed per token — input, output, and cache read are three separate prices per 1M tokens. Reasoning and thinking tokens count as output.
Cache reads are cheaper than ordinary input.
MirAPI draws from a prepaid balance: no subscription, no minimum, and the balance never expires.
When the balance runs out, requests fail with 403; topping up restores the same key immediately.

Error handling

404 on /v1/messages/count_tokens — that endpoint is not implemented; Claude Code may call it at startup, and the error is harmless.
401 — the key is missing, malformed, or unknown. Fix ANTHROPIC_AUTH_TOKEN.
403 — the key is valid but the request was refused: balance exhausted, model not whitelisted, quota reached, or IP not allowed. Check the error message for the reason.
429 — rate limited. Retry with jittered exponential backoff (1s → 2s → 4s, capped around 30s). There is no Retry-After header, and the gateway does not retry for you.
500 — safe to retry idempotent requests.
Every error message ends with a request ID — include it when you open a support ticket.

Pitfalls

Hook commands that print a lot of output slow the session down; keep them quiet.
permissions.deny blocks matching commands but is not a hard security boundary — review the interactive prompts before approving anything risky.
The Anthropic base URL stays bare (https://api.mirapi.ai); do not append /v1.