Editors & coding agents

Use Conduix from Cursor, VS Code, and OpenAI-compatible coding agents — one base URL, one key, any model in the catalog.

Every editor and coding agent on this page already knows how to talk to "a custom OpenAI-compatible endpoint." Because Conduix speaks the OpenAI Chat Completions wire format, wiring one up is configuration, not code — you point the tool at Conduix and it routes across every provider in the catalog, with your spend caps, audit trail, and PII policy applied on the way through.

The three things every tool asks for

Base URL: https://api.conduix.ai/v1
API key: cx_live_… from /dashboard/keys (use cx_test_… to trial free)
Model id: Provider-native, no prefix — e.g. claude-sonnet-5. Full list: Models

Include the /v1 in the base URL. Most tools expect the endpoint up to (but not including) /chat/completions. A few append /v1 themselves — if a tool does, drop it and use https://api.conduix.ai.

Cursor

Cursor has a built-in override for the OpenAI base URL:

Open Cursor Settings → Models.
Under OpenAI API Key, paste a cx_live_… key.
Enable Override OpenAI Base URL and set it to https://api.conduix.ai/v1.
Add the model ids you want (e.g. claude-sonnet-5, gpt-5.4) with Add model, then Verify.

Setting a custom base URL routes Cursor's Chat / Composer / Agent requests through Conduix and disables Cursor's own hosted model list. Tab autocomplete keeps using Cursor's models — that's expected. Cursor's Agent needs tool-calling, so pick a Conduix model whose catalog entry supports tools.

VS Code

VS Code has no single native AI provider — it depends on your extension. Every popular one accepts an OpenAI-compatible endpoint:

Continue.dev: Add a model to config.yaml (or config.json) with provider: openai + apiBase
Cline / Roo Code: Provider = “OpenAI Compatible” → Base URL + API Key + Model ID
GitHub Copilot: “Manage Models” BYO-key (OpenAI-compatible) — chat only, the most restricted option

Continue — config.yaml

models:
  - name: Conduix — Claude Sonnet 5
    provider: openai
    model: claude-sonnet-5
    apiBase: https://api.conduix.ai/v1
    apiKey: cx_live_YOUR_KEY

provider is the wire format, not the model vendor. That's why the example pairs provider: openai with a Claude model: openai tells the tool to speak the OpenAI Chat Completions format to apiBase, which is what Conduix accepts for every model. Conduix reads the model id and routes claude-sonnet-5 to Anthropic itself. provider: anthropic would switch the tool to Anthropic's Messages format, which the Conduix endpoint doesn't serve.

Cline and Roo Code are pure UI: pick OpenAI Compatible as the provider, then fill Base URL, API Key, and Model ID — no config file needed.

Coding agents (OpenClaw, Hermes, and friends)

Anything built on the OpenAI SDK reads a base URL and key from config or environment. For most agents, two environment variables are the whole integration:

bash

export OPENAI_BASE_URL="https://api.conduix.ai/v1"
export OPENAI_API_KEY="cx_live_YOUR_KEY"

OpenClaw: Add Conduix as a custom OpenAI-compatible provider (base_url + key) in its provider/model config; its local gateway then routes through Conduix.
Hermes Agent: Set OPENAI_BASE_URL, or Settings → Providers → custom OpenAI-compatible provider; switch models with `hermes model`.
Aider / others: The same OPENAI_BASE_URL + OPENAI_API_KEY env vars, or their --openai-api-base / --openai-api-key flags.

Model ids

Send the provider-native id — exactly what the upstream provider calls it, with no provider/model prefix. You call claude-sonnet-5 or gpt-5.4-mini, and Conduix routes to the right provider from the id alone. The live list is GET /v1/models — see Models & pricing.

Trial it without being billed. Mint a cx_test_… key and drop it into any tool above — it runs the full routing and governance path but never charges credits. Swap to a cx_live_… key when you're ready.

Govern what each editor can do

Mint one scoped key per editor, teammate, or agent rather than sharing a single key. Each key carries its own limits, so an experiment in an editor can't drain the whole balance or reach an expensive model:

allowed_models: Allowlist which models that key can call
rate_limit: Requests per minute, per key
pii_mode: Off / redact / tokenize before anything reaches a provider

Full details in Authentication and Governance.

FAQ

Base URL with or without /v1?

Use https://api.conduix.ai/v1. That's the endpoint up to /chat/completions, which is what Cursor, Continue, Cline, and most agents expect. If a tool documents that it appends /v1 for you, give it https://api.conduix.ai instead.

Which model id do I type?

The provider-native id — claude-sonnet-5, gpt-5.4, gemini-3.1-pro-preview, etc. No provider/ prefix. Fetch the current catalog from GET /v1/models.

Cursor won't verify my key, or some features are missing?

A custom OpenAI base URL disables Cursor's built-in models and leaves Tab autocomplete on Cursor's own model — that's expected. Make sure you added at least one model id and that it exists in /v1/models. For Agent mode, choose a model whose catalog entry supports tool-calling.

Can I stop a teammate's editor from burning credits on the priciest model?

Yes — give them a key with an allowed_models allowlist and a rate_limit. The key simply can't call a model that isn't on its list. Note that the credit balance and spend caps are account-level today, not per key.

Do streaming and tool calls work?

Yes. Conduix streams standard SSE (stream: true) and proxies tool/function calls, which is what agentic editors rely on. If an agent behaves oddly, it's almost always the model choice — pick one whose /v1/models entry advertises tool support.

Building on Conduix