e-INFRA CZ LLM — Claude Code Integration¶

Country-specific: This guide is for Czech academic researchers only. Access requires Metacentrum / e-INFRA CZ membership. See https://metavo.metacentrum.cz for eligibility.

What is e-INFRA CZ¶

e-INFRA CZ is the Czech national research e-infrastructure providing free LLM API access to the academic community. The LLM gateway at https://llm.ai.e-infra.cz speaks the Anthropic protocol natively (it also exposes an OpenAI-compatible API at /v1).

This changes everything about the setup: Claude Code connects directly — no proxy, no translation layer. The old Python proxy workflow is kept below as a legacy appendix only.

Get an API key at https://chat.ai.e-infra.cz → Settings → Account → API keys.

Recommended approach — native direct connection¶

Launch Claude Code with per-process environment variables. Nothing global changes, and your normal Anthropic-subscription claude keeps working in other terminals.

Unix:

CLAUDE_CONFIG_DIR="$HOME/.claude-meta" \
ANTHROPIC_BASE_URL="https://llm.ai.e-infra.cz/v1" \
ANTHROPIC_AUTH_TOKEN="<YOUR_API_KEY>" \
ANTHROPIC_MODEL="agentic" \
ANTHROPIC_SMALL_FAST_MODEL="mini" \
ANTHROPIC_DEFAULT_HAIKU_MODEL="mini" \
ANTHROPIC_DEFAULT_SONNET_MODEL="agentic" \
ANTHROPIC_DEFAULT_OPUS_MODEL="agentic" \
CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 \
claude

Windows PowerShell:

$env:CLAUDE_CONFIG_DIR = "$env:USERPROFILE\.claude-meta"
$env:ANTHROPIC_BASE_URL = "https://llm.ai.e-infra.cz/v1"
$env:ANTHROPIC_AUTH_TOKEN = "<YOUR_API_KEY>"
$env:ANTHROPIC_MODEL = "agentic"
$env:ANTHROPIC_SMALL_FAST_MODEL = "mini"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "mini"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL = "agentic"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "agentic"
$env:CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = "1"
claude

What each variable does¶

Variable	Why
`ANTHROPIC_BASE_URL`	The gateway API endpoint: `https://llm.ai.e-infra.cz/v1`
`ANTHROPIC_AUTH_TOKEN`	Your e-INFRA key. Takes precedence over subscription OAuth — this process can only talk to the gateway
`ANTHROPIC_MODEL`	The gateway model to use (see live list below)
`ANTHROPIC_SMALL_FAST_MODEL` + `ANTHROPIC_DEFAULT_HAIKU_MODEL`	Route Claude Code's background/light tasks to the gateway's small `mini` model
`ANTHROPIC_DEFAULT_SONNET_MODEL` + `ANTHROPIC_DEFAULT_OPUS_MODEL`	Map the Anthropic-named picker slots to your gateway model, so no request ever goes out with a `claude-*` id the gateway doesn't know
`CLAUDE_CONFIG_DIR`	Isolated config/credentials/history — lets subscription `claude` and gateway Claude Code run concurrently with zero shared state
`CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1`	Gateway models appear in the `/model` picker inside the session
`CLAUDE_CODE_AUTO_COMPACT_WINDOW`	Set only when the model's real context window is below Claude Code's 200k assumption (e.g. to ~75% of the real window) so auto-compaction triggers in time

Inside the session, /model <id> switches to any gateway model.

A small launcher script that wraps this — fetches the live model list, shows a picker, sets the env block, and execs claude — is the comfortable long-term setup. Store the API key in a file under the isolated config dir with chmod 600, never in a repo.

Models¶

The gateway's model list changes over time — always fetch it live rather than trusting any static table:

curl -s -H "Authorization: Bearer <YOUR_API_KEY>" https://llm.ai.e-infra.cz/v1/models

As of August 2026 the list includes agentic, coder, thinker, mini, kimi-k3, qwen3.5-122b, glm-5, deepseek, mistral-medium-3.5, gemma4. Guidance:

Agentic/tool use (Claude Code, neuroflow): agentic is the purpose-built alias; coder for coding-heavy sessions.
Vision-capable: agentic, coder, qwen3.5-122b, mistral-medium-3.5, gemma4. Text-only: glm-5, deepseek, thinker.
Background tasks: mini.
Filter out non-chat entries (embed*, rerank*, whisper*, …) when listing.

Real context window per model is available at /v1/model/info (model_info.max_input_tokens) — use it to set CLAUDE_CODE_AUTO_COMPACT_WINDOW for models under 200k.

Rate limit — the one real constraint¶

The gateway currently allows 4 parallel requests per key. Normal interactive use fits (main model + background mini ≈ 2–3 concurrent). Launching many parallel subagents will hit 429s — Claude Code retries automatically, but expect slowdowns. Prefer sequential patterns; this is why neuroflow's scholar agent searches sequentially.

Saving to neuroflow integrations.json¶

The /neuroflow:setup custom-LLM step saves the following to the chosen scope's integrations.json (global ~/.neuroflow/integrations.json or per-project .neuroflow/integrations.json):

"custom_llm": {
  "provider": "einfra",
  "base_url": "https://llm.ai.e-infra.cz/v1",
  "api_key": "<YOUR_API_KEY>",
  "model": "agentic"
}

api_key is stored locally and gitignored — never synced anywhere.
Non-secret fields can optionally be synced to your flowie private GitHub repo for cross-machine portability.
The legacy proxy_mode / proxy_port fields are only relevant to the legacy proxy appendix below.

Unrelated known issues¶

`semantic search failed: sema*...sts`¶

The Semantic Scholar MCP tool's API key is expired or rate-limited — unrelated to the gateway. Get a free key at https://www.semanticscholar.org/product/api. Workaround: use search_pubmed or search_biorxiv (no API key required).

Legacy appendix — OpenAI-compat proxy (pre-native-gateway)¶

Only needed if the native Anthropic endpoint is unavailable to you (or for other providers that are OpenAI-compatible only). For e-INFRA today, use the native approach above.

The custom FastAPI proxy (skills/setup/scripts/einfra/proxy.py) translates Anthropic↔OpenAI formats including streaming, multi-turn tool use, and thinking block passthrough. It was built because LiteLLM failed on two blocking cases with these models (Content block is not a text block with kimi thinking blocks; No tool calls but found tool output with deepseek multi-turn tool pairing).

Terminal 1 — start the proxy:

OPENAI_API_KEY=<YOUR_API_KEY> \
OPENAI_BASE_URL=https://llm.ai.e-infra.cz/v1 \
BIG_MODEL=<model> SMALL_MODEL=<model> \
uv run --python 3.12 --with fastapi --with httpx --with uvicorn \
  uvicorn proxy:app --host 0.0.0.0 --port 4001

Terminal 2 — connect Claude Code:

ANTHROPIC_BASE_URL=http://localhost:4001 ANTHROPIC_AUTH_TOKEN=dummy claude

Proxy-specific troubleshooting (Windows): port in use → netstat -ano | findstr :4001, kill via cmd.exe /c "taskkill /F /PID <PID>"; [WinError 10048] right after a kill is a TIME_WAIT ghost — switch ports instead of waiting; use port 4001+ (4000 is commonly taken). API Error: Content block not found was a proxy bug, fixed by assigning tool block indices once at creation (tool_block_started as dict, not set).

Quick reference¶

Goal	How
Connect Claude Code (native)	Set the env block above, run `claude`
Run alongside subscription Claude	`CLAUDE_CONFIG_DIR=~/.claude-meta` (isolated) — both work concurrently
List live models	`curl -s -H "Authorization: Bearer <key>" https://llm.ai.e-infra.cz/v1/models`
Switch model mid-session	`/model <id>`
Best model for tool use	`agentic`
Background/light tasks	`mini`
Context window info	`/v1/model/info` → `model_info.max_input_tokens`
Parallel request limit	4 per key — avoid heavy parallel subagent fan-out