Reaching AI Coding APIs from Europe
endpoints, hosts and what you can actually measure
What is documented is short: QCode publishes an Anthropic-shaped API at api.qcode.cc and an OpenAI-compatible endpoint at https://api.qcode.cc/v1, and its China-facing pages describe proxies placed in Asia-Pacific (Hong Kong / Japan). Nothing documents an EU point of presence, and no latency figure for Germany is published anywhere — so this page is about the switches you control and how to measure the rest.
Updated 2026-10-01
Four things you can actually control
The value that matters
Claude Code reads the endpoint from ANTHROPIC_BASE_URL, documented as overriding the API endpoint to route requests through a proxy or gateway.
Two endpoint shapes
Anthropic-shaped at api.qcode.cc, OpenAI-compatible at https://api.qcode.cc/v1 — clients that stop at /v1 append the path themselves.
Documented side effects
Claude Code's reference says MCP tool search is disabled by default when the host is not first-party, and that Remote Control is disabled when this points at a host other than api.anthropic.com.
EU point of presence
No EU PoP and no latency numbers are published on QCode's pages, so this page makes none either. The documented proxy locations are Asia-Pacific, for access from China.
What responsiveness means for a coding agent
A coding turn is not one request. The CLI sends a prompt, the model answers, a tool runs locally, its output goes back, and the loop repeats — so perceived speed is dominated by how many round trips the task needs and how long each generation takes, with the network path only one factor among them. That is why a single ping number tells you almost nothing, and why this page refuses to publish numbers it cannot source.
Why European teams ask this first
Because it is the question a relay cannot answer with a brochure. Distance to whichever host is actually serving you, session reuse, streaming behaviour, retries after a 429 or 529, and how chatty your own tool loop is — all of it lands on the same felt experience, and none of it is comparable across vendors' pages. QCode documents the endpoint values and the pricing basis; it does not publish regional timings.
Three layers, only one is yours
Office and home routing, the VPN you insist on, TLS session reuse, how many tool calls a turn makes, and whether you retry on failure.
Which host ANTHROPIC_BASE_URL or the OpenAI-compatible base URL points at, and what that host forwards to.
How long the model spends thinking and emitting — the biggest single component in an agent loop, and not a network property.
Documented vs. not published
Documented
QCode exposes api.qcode.cc for the Anthropic-shaped API and https://api.qcode.cc/v1 as an OpenAI-compatible endpoint, with callable ids on the live /models page and a price basis of official price multiplied by a per-provider service rate; its pages for access from China describe proxies deployed in Asia-Pacific (Hong Kong / Japan). Claude Code documents ANTHROPIC_BASE_URL as overriding the API endpoint to route requests through a proxy or gateway, states that MCP tool search is disabled by default when set to a non-first-party host, and that as of v2.1.196 Remote Control is disabled when this points at a host other than api.anthropic.com. Its reference also lists an environment variable to disable feature-flag fetching.
Not published — do not fill the gap
Not documented anywhere we could check: an EU point of presence for QCode, latency or throughput figures for Germany or the EU, a promised region for upstream processing, or any statement that a relay is closer to you than the first-party host. Anyone quoting millisecond numbers for this combination has not cited a source. Treat responsiveness as your own measurement, not as a claim you inherit.
Direct first-party host vs. gateway host
Direct: the vendor's own host
Keeps every first-party behaviour of the CLI intact — the Claude Code reference ties two features to the host being api.anthropic.com. This is the honest reason to prefer direct access if a feature matters to you.
Gateway: an endpoint you set yourself
ANTHROPIC_BASE_URL exists precisely to route requests through a proxy or gateway, and that is what makes shared billing and model choice possible. Documented trade-off: MCP tool search is off by default on a non-first-party host unless your proxy forwards the needed blocks, and Remote Control is disabled.
How to measure it honestly
Fix one repository and one task list, then record what the CLI already tells you: time to first visible token and total turn time, per task, at a few different hours of the day. Run the same task twice back-to-back so you can see variance. Count the tool calls inside the turn — that number, not the ping, usually explains a slow task. Compare two endpoints on the identical task list rather than on different tasks. Keep the notes with the ticket, so the decision rests on your own measurements instead of a vendor's claim; and if you disable non-essential traffic, say so in the notes, because it changes what the CLI does around the request.
On QCode
The two values you would configure are published: api.qcode.cc as ANTHROPIC_BASE_URL for the Anthropic-shaped API, and https://api.qcode.cc/v1 as the OpenAI-compatible base URL for clients that append the path themselves. Model ids come from the live /models list rather than from a page like this one, and the price basis is stated as official price multiplied by a service rate set per provider.
FAQ
Does routing through a relay change the latency?
Nobody can say from published material: QCode publishes no latency figures and no EU point of presence, and its documented proxy locations are Asia-Pacific for access from China. Treat any such claim as unverified and measure it on your own task list.
What changes when I set ANTHROPIC_BASE_URL?
Per Claude Code's own reference, the variable overrides the API endpoint to route requests through a proxy or gateway. Two behaviours are tied to the host: MCP tool search is disabled by default on a non-first-party host, and Remote Control is disabled when it points somewhere other than api.anthropic.com.
Should I use /v1 or the full path?
Stop at /v1 for OpenAI-compatible clients — QCode's own editor guides say the SDK appends the rest and warn against pasting the completions path into the URL field.
Does streaming make it feel faster?
It changes what you see first, not how long the turn takes. For an agent loop the honest metric is total turn time; time-to-first-token is worth recording alongside it, not instead of it.
What about retries after an error?
Retries are the part most teams forget, and they multiply both cost and wait time. QCode's own error pages document the quota and overload conditions to expect, which is where to look before blaming the route.
Can I cut the non-essential calls?
Claude Code's reference lists an environment variable to disable feature-flag fetching, which is a documented way to stop the CLI making extra background requests. It is a switch you should note in your measurements, not a speed promise.
Sources
Claude Code's environment-variable reference and settings documentation, fetched 2026-10-01, for the endpoint variable and the two host-dependent behaviours. QCode's own pages for the published endpoint values, the editor wiring guides, the /models catalogue, the pricing note and the pages describing proxies in Asia-Pacific for access from China (checked 2026-10-01). Where nothing is published — EU presence, regional timings — this page states that instead of estimating.
Decide with your own numbers
Two endpoints, one task list, a week of logs — then compare, instead of trusting a brochure.
Related reading
Anthropic base URL setup guide
The variable, the settings scopes, the checklist.
How to choose an AI API relay
The procurement questions worth asking a relay.
QCode pricing guide
Service rate, balance and how billing works.
No latency, throughput or availability figure is claimed on this page, because none is published for these combinations. Documented vendor behaviour was read from official pages on 2026-10-01 and can change; measure on your own infrastructure.