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 service at https://llm.ai.e-infra.cz exposes an OpenAI-compatible API, which means Claude Code can use it via a proxy that translates between Anthropic and OpenAI formats.

Available models¶

Model	Category	Best for
`kimi-k2.5`	Agentic / tool use	Recommended for Claude Code — tool calling, MCP, agentic workflows
`qwen3.5-122b`	General / best overall	Long reasoning, general tasks
`deepseek-v3.2`	Coding + reasoning	Good all-rounder
`deepseek-v3.2-thinking`	Reasoning / thinking	Complex multi-step reasoning
`qwen3-coder-next`	Coding	Latest coding model
`qwen3-coder-30b`	Fast / coding	Lighter coding model
`mistral-small-4`	Fast / small	Quick tasks

For Claude Code agentic use (tool calling, MCP tools, neuroflow), use kimi-k2.5.

Mode A — Direct connection (simplest, no model selection)¶

Set two environment variables and launch Claude Code:

ANTHROPIC_BASE_URL=https://llm.ai.e-infra.cz/v1 \
ANTHROPIC_API_KEY=<YOUR_API_KEY> \
claude

Windows PowerShell:

$env:ANTHROPIC_BASE_URL = "https://llm.ai.e-infra.cz/v1"
$env:ANTHROPIC_AUTH_TOKEN = "<YOUR_API_KEY>"
claude

⚠️ Limitations: Cannot select a specific model. The e-INFRA API may return a non-Claude model field causing Claude Code to error with "unexpected model". Use Mode B for serious agentic workflows.

Mode B — Custom Python proxy (recommended)¶

The custom FastAPI proxy (proxy.py) performs the full Anthropic↔OpenAI translation including streaming, multi-turn tool use, and thinking block passthrough. It is the most reliable option for Claude Code with e-INFRA — specifically tested and fixed for the edge cases that third-party proxies (LiteLLM) fail on with these models.

The proxy is available in skills/setup/scripts/einfra/proxy.py.

Check available models first¶

The model list at e-INFRA changes frequently. Before picking a model, fetch the current list:

curl -s -H "Authorization: Bearer <YOUR_API_KEY>" https://llm.ai.e-infra.cz/v1/models | python3 -c "import sys,json; [print(m['id']) for m in json.load(sys.stdin).get('data',[])]"

Or just ask Claude Code to do it — paste your API key and run:

fetch the model list from https://llm.ai.e-infra.cz/v1/models with my key <YOUR_API_KEY>

The table above reflects models available as of April 2026 — verify if more than a few weeks have passed.

Step-by-step workflow¶

Terminal 1 — start the proxy:

# Unix
OPENAI_API_KEY=<YOUR_API_KEY> \
OPENAI_BASE_URL=https://llm.ai.e-infra.cz/v1 \
BIG_MODEL=kimi-k2.5 \
SMALL_MODEL=kimi-k2.5 \
uv run --python 3.12 --with fastapi --with httpx --with uvicorn \
  uvicorn proxy:app --host 0.0.0.0 --port 4001

# Windows PowerShell
$env:OPENAI_API_KEY = "<YOUR_API_KEY>"
$env:OPENAI_BASE_URL = "https://llm.ai.e-infra.cz/v1"
$env:BIG_MODEL = "kimi-k2.5"
$env:SMALL_MODEL = "kimi-k2.5"
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:

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

# Windows PowerShell
$env:ANTHROPIC_BASE_URL = "http://localhost:4001"
$env:ANTHROPIC_AUTH_TOKEN = "dummy"
claude

Note: use ANTHROPIC_AUTH_TOKEN on Windows — more reliably picked up by Claude Code CLI than ANTHROPIC_API_KEY.

Switch model¶

Change BIG_MODEL and SMALL_MODEL to any model from the table above:

$env:BIG_MODEL = "deepseek-v3.2"
$env:SMALL_MODEL = "deepseek-v3.2"

What the proxy does¶

Maps all claude-* model names to the configured BIG_MODEL / SMALL_MODEL
Translates Anthropic message format → OpenAI format (including tool_use, tool_result, multi-turn)
Translates OpenAI streaming SSE → Anthropic streaming SSE with correct block indices
Restores the original claude-* model name in all responses so Claude Code doesn't reject them
Injects behavioral directives into the system prompt for better agentic performance
Passes through reasoning_content as thinking blocks

Why not LiteLLM?¶

LiteLLM was tested with e-INFRA models and produced two blocking errors:

API Error: Content block is not a text block — kimi-k2.5 returns thinking/reasoning blocks that LiteLLM passes through incorrectly, causing Claude Code to find a thinking block where it expects text.
No tool calls but found tool output — deepseek-v3.2 rejects multi-turn conversations where LiteLLM's tool result→tool message translation loses the pairing with the preceding tool call.

The custom proxy.py handles both cases correctly. LiteLLM may work in a future version — but as of April 2026, use proxy.py for e-INFRA.

Port conflict resolution (Windows)¶

A common Windows issue: port already in use ([WinError 10048]).

Check what's using the port:

netstat -ano | findstr :4001

Kill the process (must use cmd.exe — taskkill breaks in Git Bash on Windows):

cmd.exe /c "taskkill /F /PID <PID>"

If the process doesn't appear in tasklist but netstat still shows it — ghost socket in TIME_WAIT. Switch to a different port rather than waiting:

# change --port 4001 to --port 4002

Use port 4001 or higher — port 4000 is commonly grabbed by other services on Windows.

Known issues and fixes¶

`API Error: Content block not found`¶

What it means: Claude Code's streaming parser received a content_block_delta or content_block_stop referencing an index never opened with content_block_start.

Root cause in proxy.py (fixed): tool_block_started was a set tracking which tool indices had started, but not which Anthropic block index they were assigned. The final cleanup recalculated next_idx + idx which was stale.

Fix: tool_block_started is now a dict mapping openai_tool_idx → assigned_block_index. Block indices are assigned once at creation time and stored — never recalculated.

`API Error: Content block is not a text block`¶

LiteLLM + kimi-k2.5 issue — kimi returns thinking/reasoning blocks that LiteLLM passes through as Anthropic thinking blocks. Claude Code finds a thinking block where it expects text. Use proxy.py instead.

`No tool calls but found tool output`¶

LiteLLM + deepseek-v3.2 issue — message history translation loses the pairing between tool calls and tool results in multi-turn conversations. Use proxy.py instead.

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

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

Claude Code freezes or hangs with many concurrent tool calls¶

e-INFRA models can time out under high concurrency. neuroflow's scholar agent runs searches sequentially for this reason. For other workflows: prefer sequential tool use patterns.

`[WinError 10048]` immediately after killing a process¶

Windows TIME_WAIT state — port locked for up to 60 seconds after process death. Switch port instead of waiting.

Saving to neuroflow integrations.json¶

The /neuroflow:setup Step 5 saves the following to .neuroflow/integrations.json:

"custom_llm": {
  "provider": "einfra",
  "base_url": "https://llm.ai.e-infra.cz/v1",
  "api_key": "<YOUR_API_KEY>",
  "model": "kimi-k2.5",
  "proxy_mode": "custom",
  "proxy_port": 4001
}

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.

Quick reference¶

Goal	Command
Start proxy (Unix)	`BIG_MODEL=kimi-k2.5 OPENAI_API_KEY=<key> uv run ... uvicorn proxy:app --port 4001`
Start proxy (Windows)	`.\start_proxy.ps1`
Connect Claude Code (Unix)	`ANTHROPIC_BASE_URL=http://localhost:4001 ANTHROPIC_AUTH_TOKEN=dummy claude`
Connect Claude Code (Windows)	`$env:ANTHROPIC_BASE_URL="http://localhost:4001"; $env:ANTHROPIC_AUTH_TOKEN="dummy"; claude`
Check port in use (Windows)	`netstat -ano \\| findstr :4001`
Kill port owner (Windows)	`cmd.exe /c "taskkill /F /PID <PID>"`
Switch model	Set `BIG_MODEL=<model>` env var before starting
Best model for tool use	`kimi-k2.5`
Best model for reasoning	`deepseek-v3.2-thinking`