Every tool that speaks OpenAI, pointed at one endpoint

By the aitokensai editors · Verified 2026-11-09 Sources: OpenAI API docs

Nearly every AI coding tool and framework ships one assumption: you have an OpenAI key and you talk to OpenAI. Almost all of them also expose the one setting that breaks that assumption. Change it once and every model becomes reachable. Without waiting for tool support to add the vendor you actually want to pay for.

The three values that decide everything

Every integration on this page is the same three values wearing different clothes. Learn them once and the per-tool instructions become trivial:

curl https://aicomp.ai/v1/models \
  -H "Authorization: Bearer sk-your-gateway-key"

The one line every tool and framework needs

curl https://aicomp.ai/v1/chat/completions \
  -H "Authorization: Bearer $GATEWAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-pro","max_tokens":256,\
       "messages":[{"role":"user","content":"Reply with the word ok."}]}'

Before touching any tool config, run that command. If it returns a completion, you have proven all three values work. Every tool-level failure from this point on is a configuration problem in the tool, not an API problem.

Tool matrix

Menu labels move between versions; the configuration key or environment variable does not. Where the two disagree in our guides, trust the variable.

The three failures behind most support tickets

What it costs to run an agent all day

Coding agents are output-heavy: short prompts in, long diffs out. That is why the output column matters more than the input column everyone compares. The table below prices one heavy day of agent work, 200k input, 60k output, across twelve models that people actually route to from these tools.

When it is configured but still not working

Every page above ends with a working setup. This is where to go when one does not — the ten diagnostics below cover the failures that show up after the configuration looks correct, which is where most of the time actually goes.

The shared idea across all ten: a URL that is never read and a URL that is wrong look identical, and authentication happens before routing So a 401 localises to the credential and a 404 to the path. The same logic runs one layer down: a name that routes and a name that resolves are different things, and only one of them tells you when it failed.

Change models without touching credentials

The reason to keep one endpoint is that switching models stops being a project. Nothing about the key or the URL changes — only the string you send:

# Same key, same URL, different partner:
# change this one line and the entire bill changes shape
export MODEL=deepseek-v4-flash

curl "$GATEWAY_BASE_URL/chat/completions" \
  -H "Authorization: Bearer $GATEWAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{{\"model\":\"$MODEL\",\"max_tokens\":128,\
       \"messages\":[{{\"role\":\"user\",\"content\":\"ok?\"}}]}}"

Put that environment variable somewhere every tool reads from and you can promote a failing call to a stronger model, or demote an expensive one to a cheaper tier, without opening a single settings panel. In practice this is how teams end up running three models at once: a strong one for anything touching production code, a mid-priced one for the daily grind, and an inexpensive one for summaries and autocomplete.

What "OpenAI-compatible" does not promise

The phrase is narrower than it sounds. It describes a request shape, not a guarantee that every feature survives the trip:

Read that as: switching models through one endpoint is cheap, but it is not free. Budget ten minutes per new model to re-check tool reliability and cost behaviour rather than assuming the previous model's results transfer.

Rollout checklist

• Prove it with curl first. Every tool-level failure is cheaper to diagnose outside the tool.
• Put the three values in the environment. Base URL, key, model ID (never inline them in a file that gets committed).
• Confirm in your usage log, not in the UI. A green checkmark means a request succeeded, not that it billed to you.
• Pin the model ID explicitly. Anything defaulting to a snapshot will silently change behaviour when your endpoint's catalogue updates.
• Re-run your own evaluation prompt after switching. Tokenisation differs, so yesterday's cost per task is not today's.

FAQ

Related

• LiteLLM: one interface, per-model api_base, and fallbacks
• Vercel AI SDK with a custom endpoint
• Cline with an OpenAI-compatible provider
• Claude Code against your own key
• Cursor with an overridden base URL
• Cursor Agent mode: which surfaces reach your endpoint
• Cursor Agent stalls while chat answers
• What OpenAI-compatible actually guarantees
• No cheapest model — only a cheapest token mix
• Full price index
• What one endpoint makes possible: a fallback chain in one config

Get API access