Cline with an OpenAI-compatible endpoint
Cline assumes OpenAI by default, but it ships a generic OpenAI-compatible provider whose only visible difference is one extra field. Fill it in and the extension stops caring which company serves the request — which also means you can move expensive tasks onto a cheaper model without reinstalling anything.
What you need
- A gateway API key.
- The base URL:
https://aicomp.ai/v1. - An exact model ID from
GET /models.
curl https://aicomp.ai/v1/models \
-H "Authorization: Bearer sk-your-gateway-key"A JSON list of model IDs means the key is good. Invalid token means it was copied wrong.
The setting that matters
| Field | Value | Notes |
|---|---|---|
| API Provider | OpenAI Compatible | This is what reveals the base URL field at all. |
| Base URL | https://aicomp.ai/v1 | Ends at /v1. The client appends the rest. |
| API Key | Your gateway key | Paste, do not retype — trailing spaces produce 401s. |
| Model ID | Exact string | Typed by hand if absent from the dropdown. |
Menu labels differ between versions; field semantics do not. If your build words the provider differently, it is the one that unlocks a base URL.
Keep the values out of your editor state
Never paste either value into a project file — they end up in git history and then in a public repository. Export them in the shell you launch your editor from, so nothing sensitive lives in the workspace:
# In the shell you launch your editor from — not in a committed file
export GATEWAY_BASE_URL="https://aicomp.ai/v1"
export GATEWAY_API_KEY="sk-your-gateway-key"
# List what you can actually call, and copy IDs from here rather than typing them
curl -s "$GATEWAY_BASE_URL/models" \
-H "Authorization: Bearer $GATEWAY_API_KEY" | jq -r '.data[].id' | head -40Copy model IDs from that list rather than reconstructing them from memory. Vendor APIs reject unknown model strings outright — there is no fuzzy matching and no fallback to a similar model, which is why a typo looks like an outage.
Why the base URL must end at /v1
The field is a prefix, not a full address. Whatever you type, the client appends the resource path before sending — so the value has to stop before it:
# Correct — client appends /chat/completions to this prefix
https://aicomp.ai/v1 -> https://aicomp.ai/v1/chat/completions ✅
# Wrong — the path is now doubled, and every request 404s
https://aicomp.ai/v1/chat/completions
-> https://aicomp.ai/v1/chat/completions/chat/completions ❌
# Wrong in the other direction — silently talks to OpenAI instead
(empty field) -> https://api.openai.com/v1/... ⚠️The third case is the expensive one: it works. Requests succeed, tasks complete, and the bill lands somewhere you were not watching. This is exactly why confirming in your gateway usage log matters more than seeing a successful task.
Plan and Act burn tokens differently
Most agent extensions separate planning from acting, and that split is the cheapest control you have:
- Planning reads a lot. It gathers context, so it is input-heavy — the rate that everyone quotes and compares.
- Acting writes a lot. It produces patches, explanations and retries, so it is output-heavy — usually four times the input rate at median across the catalogue.
- Therefore: model choice per phase beats one model for everything. A stronger model for planning decisions and an inexpensive one for mechanical edits usually lands below running everything on the strongest row.
Why the agent workload changes the maths
An agent is not a chatbot. Each task reads files, writes a plan, edits, runs a command, reads the error and tries again — which is a mountain of output tokens relative to input. That is why the output rate decides the bill while the input rate everyone compares barely registers:
| Model | Vendor | Rate in / out per 1M | Per day | Per month |
|---|---|---|---|---|
| gpt-5.6-luna | OpenAI | $0.1 / $0.6 | $0.06 | $1 |
| MiniMax-M3 | MiniMax | $0.15 / $0.6 | $0.07 | $1 |
| deepseek-v4-flash | DeepSeek | $0.22 / $0.66 | $0.08 | $2 |
| gemini-3.7-flash | $0.375 / $1.875 | $0.19 | $4 | |
| claude-haiku-4-5-20251001 | Anthropic | $0.5 / $2.5 | $0.25 | $5 |
| deepseek-v4-pro | DeepSeek | $0.66 / $1.98 | $0.25 | $5 |
| glm-5.3 | Zhipu | $0.7 / $2.2 | $0.27 | $5 |
| qwen3.8-max | Alibaba | $1 / $3 | $0.38 | $8 |
| claude-sonnet-5 | Anthropic | $1 / $5 | $0.50 | $10 |
| gpt-5.6-terra | OpenAI | $1 / $6 | $0.56 | $11 |
| kimi-k3 | Moonshot | $1.5 / $7.5 | $0.75 | $15 |
| claude-opus-5 | Anthropic | $2.5 / $12.5 | $1.25 | $25 |
https://aicomp.ai/v1).
Create one free →
How this fails in practice
| Symptom | Cause | Fix |
|---|---|---|
| 404 on every request | Doubled path from adding /chat/completions yourself | Base URL ends at /v1 |
| No Base URL field visible | Still on a named vendor provider | Switch provider to OpenAI Compatible |
| 401 immediately | Whitespace in a hand-typed key | Re-paste; check for a stray newline |
| Works but bills the wrong account | Empty field falling back to OpenAI | Confirm the request appears in your gateway log |
| Model not selectable | Static dropdown, not a live catalogue | Type the ID from GET /models |
FAQ
Where is the Base URL field?
It appears once the provider is set to OpenAI Compatible. Under a named provider the field is not shown, which is why people conclude the feature does not exist. Menu wording moves between versions — the provider choice is what matters.
My first task succeeded, so why is nothing in my usage log?
Either the override was not saved or the editor was not restarted. Some settings are read once at startup; a reload is enough in most builds, a full quit is sometimes required. Absence from your log is the only reliable signal.
Every request 404s.
Almost always a doubled path. The client appends the resource path to your base URL, so it must end at /v1. The opposite failure — leaving the field empty — sends requests to OpenAI instead, which is worse because it looks like it works.
My model is not in the list.
Type it. The list is a snapshot shipped with the extension, while your endpoint's catalogue changes whenever upstream does. The exact ID from GET /models is what gets billed.
Can I use one key across several tools?
Yes, and that is the practical reason to route through a gateway — one key, one bill, and spend visible per model instead of per vendor dashboard. Keep it out of any file that gets committed.