Claude Code and Codex CLI on a pay-as-you-go API key

1. Create a key per developer

   In API keys, create one key for each person or CI job and give it a spend limit. Use in tools on any key opens ready-made setups for each CLI, filled in with the key.

2. Point your agent at the gateway

   Set the base URL and the key in your shell or the agent's config file. Each CLI talks the API it was built for: Anthropic, OpenAI or Gemini.

3. Code as usual

   Every session is pinned to one channel so its prompt cache stays warm. Logs show each call's tokens, cache hits and charge, per key.

Claude Code#

Export the settings and start it, or put them in ~/.claude/settings.json to keep them across sessions:

export ANTHROPIC_BASE_URL=https://lovelyrouter.com
export ANTHROPIC_AUTH_TOKEN=$LOVELY_API_KEY
export ANTHROPIC_MODEL=claude-sonnet-5.5
claude

ANTHROPIC_MODEL is the main model and ANTHROPIC_DEFAULT_HAIKU_MODEL handles quick background tasks; any Claude model id works, such as claude-opus-5.5. See the Claude Code guide and the Claude API page.

Codex CLI#

Codex CLI uses the OpenAI Responses API. Add Lovely Router as a model provider, then export the key named in env_key:

model = "gpt-6.1-sol"
model_provider = "lovely"

[model_providers.lovely]
name = "Lovely Router"
base_url = "https://lovelyrouter.com/v1"
env_key = "LOVELY_API_KEY"
wire_api = "responses"

export LOVELY_API_KEY="lr-..."
codex

Use a GPT model (gpt-*): those are the models served over the Responses API. Current Codex versions no longer accept wire_api = "chat". See the Codex CLI guide and GPT models.

Gemini CLI#

export GOOGLE_GEMINI_BASE_URL=https://lovelyrouter.com
export GEMINI_API_KEY=$LOVELY_API_KEY
gemini -m gemini-3.1-pro

A coding agent resends the whole conversation on every turn: the system prompt, the tool definitions and every file it has read so far. With prompt caching, most of those input tokens are billed at the cache-read price, a fraction of the normal input price. But a cache lives on the channel that wrote it. If a turn lands on another channel, the cache isn't there, and the whole prefix is paid for again.

The gateway keeps each conversation on the channel that served its first turn:

Each call in Logs shows its cached tokens, cache hit rate and affinity (sticky, new or explicit). Prompt caching has the details.

All keys draw on one balance, but each has its own limits, and the key list shows what each one has spent.

A key that reaches its spend limit gets 402 with This API key has reached its spend limit.; a rate or concurrency limit returns 429 with a Retry-After header. Refused requests are not billed. See API keys.

The same key connects the MCP server, so the agent can search the web, read a docs page as Markdown, pull structured data from Google Maps, LinkedIn or Reddit, and drive a browser through a multi-step flow:

claude mcp add --transport http lovely-tools https://lovelyrouter.com/mcp \
  --header "Authorization: Bearer $LOVELY_API_KEY"

Web search covers lookups, web fetch reads one page and website crawl a whole docs section. The browser agent works in the cloud, so it handles public sites, not localhost.

Model calls are billed per token: input, output, cache reads and cache writes, each at the model's price per million tokens. Requests that return an error are not charged. Data tools are billed per result and the browser agent per step, from the same balance. There is no subscription and no automatic recharge: you top up from $5 and are never billed more than you added. See pricing and the full model list.