# Claude Code

> Point Claude Code at Tchavi with two environment variables and run it on credits topped up by mobile money.

- Canonical: https://tchavi.com/en/docs/coding-agents/claude-code

---


Part of the [coding agents](/docs/coding-agents) guides.

Claude Code speaks the Anthropic Messages API and nothing else — it cannot be pointed at an OpenAI-compatible endpoint. Tchavi therefore serves it a dedicated `/v1/messages` endpoint. Everything else — credits, rate limits, request logs — works exactly as it does for the rest of the API.

## 1. Set the variables

```bash
export ANTHROPIC_BASE_URL=https://tchavi.com/api
export ANTHROPIC_AUTH_TOKEN=sk-tch-your-key
```

Note the base URL stops at `/api`: Claude Code appends `/v1/messages` itself.

Use `ANTHROPIC_AUTH_TOKEN`, not `ANTHROPIC_API_KEY`. Both work — Tchavi reads the key from either the `Authorization` or the `x-api-key` header — but `ANTHROPIC_API_KEY` needs a one-time approval inside Claude Code before it takes effect, and it is the more common source of a session that silently keeps using your old login.

If you were already signed in to claude.ai, run `/logout` once so the two credentials stop competing.

## 2. Check it before opening the agent

One request tells you whether the URL and the key are right, which is much easier to read than a failure inside the agent:

```bash
curl -sS -w '\n%{http_code}\n' -X POST "$ANTHROPIC_BASE_URL/v1/messages" \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-sonnet-4-6","max_tokens":16,"messages":[{"role":"user","content":"Say OK"}]}'
```

A body starting with `{"id":"msg_` and ending in `200` means you are done. A `401` means the key did not arrive — check for a typo and that you exported it in the shell you are about to run `claude` in.

## 3. Choose your models

Claude Code's built-in model names are Anthropic's, and Tchavi carries those same names, so it works with no model configuration at all. Pin them explicitly to control what each alias costs you:

```bash
export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4-8
export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-4-6
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5-20251001
```

`ANTHROPIC_DEFAULT_HAIKU_MODEL` is worth setting deliberately: it also runs the background work — session titles, summaries — that you never see but do pay for.

You are not limited to Claude. Any tool-capable chat model in the catalogue works, because Tchavi translates between the two formats for you:

```bash
export ANTHROPIC_MODEL=gpt-5.4          # or deepseek-v4-pro, kimi-k3, qwen3.8-max
```

<Callout type="tip" title="Cheaper background work">
  Pointing the Haiku alias at a budget model — `deepseek-v4-flash` at 0.72 credits per 1K input, or
  `glm-5.3-flash` at 0.11 — cuts the cost of background tasks sharply without touching the quality
  of your actual coding turns.
</Callout>

## 4. Make it permanent

Shell exports last one terminal. To apply everywhere Claude Code runs, put them in the `env` block of `~/.claude/settings.json`:

```json
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://tchavi.com/api",
    "ANTHROPIC_AUTH_TOKEN": "sk-tch-your-key",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5-20251001"
  }
}
```

Use `~/.claude/settings.json` (yours, all projects) or `.claude/settings.local.json` (one project, git-ignored). Never a project's `.claude/settings.json` — that file is committed, and your key would go with it.

## 5. Show Tchavi models in the picker

Claude Code can read the catalogue from Tchavi and add it to `/model`:

```bash
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1
```

It calls `GET /v1/models` at startup and keeps only entries whose id contains `claude` or `anthropic`. So the Claude models appear; `gpt-5.4` and `deepseek-v4-pro` do not, and you select those with `ANTHROPIC_MODEL` instead.

## What does not carry over

These are worth knowing before you decide, rather than discovering mid-task:

- **Prompt caching is not applied to Claude models.** Claude Code marks parts of the conversation as cacheable; that marking is lost translating to and from the OpenAI shape, so a long session on a Claude model bills every turn as fresh input. On a big repository this is the difference that matters most. Point `ANTHROPIC_MODEL` at an OpenAI-backed model instead and caching does apply — the provider caches long prompts on its own, with no marker needed, and Tchavi bills those tokens at a tenth of the input rate (see [Credits](/en/docs/credits#prompt-caching)).
- **Remote Control and voice dictation stop working.** Both require a claude.ai identity, which a gateway credential replaces.
- **`/fast` reports unavailable.** Its availability check goes to Anthropic directly and does not follow your base URL.
- **`/context` is approximate.** Tchavi answers the token-counting endpoint with a character-based estimate, not a real tokenizer. It is accurate enough to steer compaction; your bill uses the real counts the provider reports.

A handful of models cap the prompt they can be billed for — `gemini-3.1-pro` and `grok-4.6` at 180,000 tokens, `qwen3.7-flash` at 230,000. Claude Code does not recognise Tchavi's refusal as a too-long error, so it will not compact and retry by itself. If you use one of those, tell it the ceiling:

```bash
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=180000
```

Claude models carry no such cap.

---

Next: [OpenCode](/docs/coding-agents/opencode) · Back to [coding agents](/docs/coding-agents)