Claude Agent SDK

Searching the SDK for a base_url option finds nothing, and that is not an omission in the docs — the option does not exist. The SDK runs the Claude Code CLI as a subprocess, and the CLI is the thing that reads ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. Set those two and every call the agent makes is routed, with no change to your agent code.

# The SDK has no base_url option. The CLI it spawns reads these, and the
# SDK passes the parent environment straight through — so exporting them
# before your program starts is enough.
export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-kn-...

python my_agent.py

Why the environment reaches the CLI at all

This is worth one paragraph, because it is the difference between a trick that might stop working and a documented property you can build on. The Python SDK's subprocess transport constructs the child's environment as the parent's os.environ with a single key removed — CLAUDECODE, so the child does not think it is running inside a Claude Code session — then merges CLAUDE_CODE_ENTRYPOINT, then ClaudeAgentOptions.env, then the SDK version.

Two things follow, and the second is the one people get wrong. Everything in your shell reaches the CLI, so exporting the two variables works. And options.env is merged on top of the inherited environment, so the explicit form wins over a stale export rather than losing to it. The code is in subprocess_cli.py.

The explicit form, and when to insist on it

Exported variables are fine on your own machine and fragile everywhere else: the agent's endpoint becomes a property of how the process was launched, which breaks the first time it runs under a scheduler, in a container, or in a CI job that does not source your profile. Passing env on the options object makes routing part of the program.

from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions

# The explicit form. options.env is merged ON TOP of the inherited
# environment, so this wins over whatever the shell happens to hold —
# which is what you want in anything that is not your own laptop.
options = ClaudeAgentOptions(
    env={
        "ANTHROPIC_BASE_URL": "https://api.kunavo.com",
        "ANTHROPIC_AUTH_TOKEN": "sk-kn-...",
    },
)

async with ClaudeSDKClient(options=options) as client:
    await client.query("Summarise the open TODOs in this repo")
    async for message in client.receive_response():
        print(message)

Step by step

1. Create a key at /app/keys and copy it — it is shown once.
2. Decide where routing lives: exported variables for local work, ClaudeAgentOptions(env=…) for anything that runs unattended.
3. Set ANTHROPIC_BASE_URL to https://api.kunavo.com and ANTHROPIC_AUTH_TOKEN to your sk-kn-… key.
4. Optionally set ANTHROPIC_DEFAULT_HAIKU_MODEL so the background sub-tasks the CLI fires land on the cheapest tier.
5. Run your program. Nothing in the agent code changes.

Which tier for which sub-task

An agent fans out — one thing you ask for becomes many billed round trips — so the tier mapping matters more here than in a chat app. Rates are USD per 1M tokens, input / output, read live from the catalog.

Verify before you debug the SDK

One request settles whether a failure is the key, the endpoint, or the SDK. If this returns 200, the same credential works for the CLI the SDK spawns, and anything still broken is where the variables are set rather than what they contain.

# Settles whether a failure is the key, the endpoint, or the SDK.
# 200 here means the same credential works for the CLI the SDK spawns.
curl -sS https://api.kunavo.com/v1/messages \
  -H "Authorization: Bearer sk-kn-..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-haiku-4-5","max_tokens":16,
       "messages":[{"role":"user","content":"ping"}]}'

Reference

The SDK is open source at anthropics/claude-agent-sdk-python. The environment behaviour described here is from its own subprocess transport, read on 2026-09-04. The TypeScript SDK has the same architecture — it drives the CLI rather than calling the API — so the CLI again reads the routing variables; its README documents neither the option nor the behaviour, so confirm the option name in its types before relying on the explicit form there. Kunavo's side is the Messages API, and the other clients that route the same way are on the integrations hub.