Configuring Cursor and Cline to work with a relay service

What makes editor tools different

Tools like Cursor, Cline, Continue and Roo Code are not used the way Claude Code is, and that produces two special requirements:

One: the requests are small and constant. Every few keystrokes, every file save, can trigger a request. That can add up to several hundred or a thousand calls a day.

Two: they are extremely sensitive to time-to-first-token. A command-line tool can afford to be slow; a completion that lags half a second in the editor breaks your flow.

Together these decide one thing: when you pick a service for editor work, latency matters more than the rate multiplier.

What to configure

1. base_url and protocol

Most editor tools speak the OpenAI-compatible protocol, which keeps configuration simple:

Base URL: https://service-domain/v1
API Key:  your key
Model:    the model id the service offers

Watch whether base_url should end in /v1 — tools handle it differently, and getting it wrong means a 404. Trying both once is the fastest way to find out.

2. Model choice

For editor work, pick a model that is fast and cheap rather than the strongest one:

• Completion, renaming, small refactors → a lightweight model is enough
• Complex refactors, architecture work → switch to a strong model by hand

Running every completion through the strongest model is absurdly expensive, and slow.

3. Turn off automatic requests you do not need

Many editor tools enable completion-as-you-type by default. That is a plus for focus and a minus for cost.

Suggested:

• Keep inline completion (it is the core of the experience)
• Turn off high-frequency whole-file analysis
• Trigger it manually when you need it

Three common traps

Trap 1: hitting the concurrency limit

Editor tools can fire several requests in a burst (completion + background analysis + indexing). If your service caps concurrency low, you get scattered failures.

Symptom: most requests are fine, one or two error out. Because it does not break the main flow it is easy to ignore — but you are quietly losing requests the whole time.

Fix: cap concurrency in the tool's settings, or move to a service with more headroom.

Trap 2: silently truncated context

Editor tools send a whole file, or a whole project slice, as context. If the service limits input length, the excess may be silently truncated rather than returned as an error.

Symptom: completions on small files are fine; on large files the output is obviously wrong, because the model only ever saw part of the file.

How to test: take a large file and put something near the end that can only be completed correctly if the whole file was read.

Trap 3: caching that does not apply, and the bill that follows

Editor work repeats context heavily (the same file, over and over), so caching should save a great deal. If the service does not support caching, you pay for the full file on every resend.

The arithmetic: a 2,000-line file is roughly 30k tokens. Billed per call, 200 completions a day is 6 million tokens. With caching and without it, that is several times the cost.

Configuration checklist

[ ] Tried base_url both with and without /v1, and know which one works
[ ] Model name matches what /v1/models returns
[ ] Tested with a large file, confirmed context is not truncated
[ ] Turned off unnecessary high-frequency automatic requests
[ ] Watched time-to-first-token, confirmed streaming works
[ ] After a day of use, checked billing to confirm caching applies
[ ] Confirmed the concurrency cap is enough

One recommendation

For editor work, use a service of its own — do not share it with production traffic.

The reason: editor traffic is heavy and bursty, and can trip risk controls that then affect your production work; in the other direction, production spikes will degrade your completions. Manage them separately and tune each on its own.

Related reading

• Choosing a provider for code completion: why latency beats rate
• Does prompt caching actually work on a relay service?
• How to give a project several API providers as failover