Claude Code

Claude Code is Anthropic's official terminal coding agent — a single claude command that reads files, makes edits, runs tools, and holds a session's worth of context across turns. The distinguishing behavior is how hard it pushes on exploration: slash commands, long tool chains, multi-file edits, thorough summaries. Because it reads ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, pointing it at Melious is two environment variables — our Messages endpoint implements the Anthropic SDK and maps claude-* model names to the open-weight models we actually run.

With the Melious CLI

With the Melious CLI installed and signed in, two commands take you from nothing to a session:

melious tools install claude-code
melious launch claude-code

For Claude Code, tools install prints Anthropic's installer command for your OS rather than running it — run that line yourself, then launch from inside your project. The first launch in a project asks Trust this workspace? The agent runs as you, with your Melious key, so Melious asks before handing it over; the answer defaults to no. Workspace trust covers what a grant means and how to revoke it.

Melious configures Claude Code through environment variables only, scoped to the launched process: the base URL, your key as ANTHROPIC_AUTH_TOKEN, the model slots (main, secondary, lightweight), and a longer request timeout. It clears an inherited ANTHROPIC_API_KEY and the other-provider selectors for that process, and writes no file — your Claude settings, sessions, and login stay where Claude Code put them. Melious sets CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1, which reduces nonessential traffic; Claude Code's own policies apply. Models, -- passthrough, and exit status: launching a coding agent.

Manual setup (alternative)

For when you'd rather not use the Melious CLI. These steps set the same variables by hand.

curl -fsSL https://claude.ai/install.sh | bash

irm https://claude.ai/install.ps1 | iex

brew install --cask claude-code            # macOS, Linux
winget install Anthropic.ClaudeCode        # Windows
npm install -g @anthropic-ai/claude-code   # any platform, needs Node.js

# Required.
export ANTHROPIC_BASE_URL=https://api.melious.ai
export ANTHROPIC_AUTH_TOKEN=sk-mel-<YOUR_API_KEY>
export ANTHROPIC_API_KEY=""

# Recommended. Pins every slot Claude Code may reach for at
# runtime — including hard-coded names like `claude-haiku-4-5`
# Anthropic uses for background tasks — to a known Melious model.
# Also raises the request timeout for long agent runs.
# CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC reduces nonessential traffic
# to Anthropic; Claude Code's own policies apply.
export ANTHROPIC_MODEL=<MODEL_ID>
export ANTHROPIC_SMALL_FAST_MODEL=<LIGHTWEIGHT_MODEL_ID>
export ANTHROPIC_DEFAULT_OPUS_MODEL=<MODEL_ID>
export ANTHROPIC_DEFAULT_SONNET_MODEL=<MODEL_ID>
export ANTHROPIC_DEFAULT_HAIKU_MODEL=<LIGHTWEIGHT_MODEL_ID>
export API_TIMEOUT_MS=3000000
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.melious.ai",
    "ANTHROPIC_AUTH_TOKEN": "sk-mel-<YOUR_API_KEY>",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_MODEL": "<MODEL_ID>",
    "ANTHROPIC_SMALL_FAST_MODEL": "<LIGHTWEIGHT_MODEL_ID>",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "<MODEL_ID>",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "<MODEL_ID>",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "<LIGHTWEIGHT_MODEL_ID>",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

claude

Auth token: ANTHROPIC_AUTH_TOKEN
Anthropic base URL: https://api.melious.ai

How model names are mapped

Claude Code hard-codes Anthropic model names (opus, sonnet, haiku). We match these names and route each tier to a Melious model so you don't need a patched client:

The response carries the original claude-sonnet-4 (or whichever name you sent) back in the model field — so Claude Code's internal checks don't trip.

Browse melious.ai/hub/models for the current model list.

Picking a specific model

Most current Claude Code releases accept arbitrary --model strings, so this works:

claude --model <MODEL_ID>

If your version validates against Anthropic's allow-list, set the ANTHROPIC_DEFAULT_*_MODEL env vars instead — Claude Code passes whatever string you put there straight through to the API:

export ANTHROPIC_MODEL=<MODEL_ID>
export ANTHROPIC_SMALL_FAST_MODEL=<LIGHTWEIGHT_MODEL_ID>
export ANTHROPIC_DEFAULT_OPUS_MODEL=<MODEL_ID>
export ANTHROPIC_DEFAULT_SONNET_MODEL=<MODEL_ID>
export ANTHROPIC_DEFAULT_HAIKU_MODEL=<LIGHTWEIGHT_MODEL_ID>
export CLAUDE_CODE_SUBAGENT_MODEL=<MODEL_ID>

Claude Code passes these through verbatim, and we run the string as a model ID — an ID from GET /v1/models works, anything else is a 404 not_found_error — so they're the most reliable way to lock in a specific model. Pin every slot — Claude Code uses *_HAIKU_* and SMALL_FAST_MODEL for background and agent-spawn calls, so leaving them unset can route quietly to whatever's mapped to haiku. Find current IDs at melious.ai/hub/models.

What's different

• No environment_impact or billing_cost in responses. Claude Code talks to our Messages endpoint, which returns neither field, streamed or not — there's nothing for the SDK to pass along. Usage is still logged and surfaced in your usage dashboard.
• Explicit cache_control blocks aren't exposed. There's no handle to mark a prefix as cacheable. Transparent prefix caching still happens on the provider side and shows up as usage.cache_read_input_tokens, billed at the cheaper cache-read rate — so repeated context within a session does benefit, even though Claude Code can't set cache breakpoints. See Models.
• Extended thinking works on reasoning models. Send thinking: {"type": "enabled", "budget_tokens": 8000} to turn it on, {"type": "disabled"} to turn it off, or leave it out for the model's default. Claude Code's effort setting reaches the model too — it's sent as output_config.effort and wins over the budget. xhigh and max arrive as sent only on glm-5.3 and glm-5.3-flash; on every other model they fold to high. Thinking comes back as a thinking content block before the text (signature is always an empty string — don't send it back expecting Anthropic to verify it).

When it breaks

• 404 on every request — ANTHROPIC_BASE_URL ends in /v1. The SDK appends /v1/messages itself, so the request goes to /v1/v1/messages, which doesn't exist. Set the base URL to https://api.melious.ai.
• 401 on every request — three causes. The token string lost its sk-mel- prefix when you copied it, or the key was rotated: paste the full 128-character key. Or ANTHROPIC_API_KEY is set somewhere (shell, dotfile, CI secret): the SDK sends it as x-api-key, which we check before the bearer token. Blank it: export ANTHROPIC_API_KEY="". If you previously ran claude login, also /logout to drop cached creds.
• /status shows the wrong base URL — env vars didn't reach the process. Re-export in the same shell you launched claude from, or move the values into .claude/settings.local.json so they're read at startup regardless of shell state.
• Tool loops feel different than with Claude — model-capability thing, not a protocol thing. Stronger models handle longer tool sequences; smaller ones give up earlier. Check _meta.capabilities.function_calling on GET /v1/models?include_meta=true and override the mapping with ANTHROPIC_DEFAULT_*_MODEL for long agent sessions.

Errors and retry patterns: Errors.