Can you run Continuum’s models in Devin?

The short answer, and the sentence it rests on

There is no setting, no environment variable, no configuration file key, and no enterprise contract term that makes Devin call an endpoint you operate. Cognition’s own Enterprise Deployment documentation puts it in a single sentence, which is worth quoting exactly because it is unusually unambiguous for a vendor FAQ:

Two things follow from that, and neither is softened anywhere else in the documentation. There is no bring-your-own-key path, and there is no custom base URL. The word "currently" is Cognition’s, not ours; it leaves the door open without committing to anything, and it would be dishonest to read it as a roadmap.

The Assured tier is the one people expect to be different, because customer-managed keys sound adjacent to bringing your own model. They are not the same thing: CMK is about encryption of data at rest in the deployment, not about where inference runs. The FAQ sentence sits on the enterprise deployment page and is not scoped to a tier.

What Devin does let you change

Devin is not modelless. It exposes a real model selector; the selector just picks from Cognition’s roster rather than pointing at an endpoint. It is worth knowing properly, because it is the whole of the control surface you get.

• Mid-session, in the CLI: /model opus, /model codex, and so on. Running /model with no argument opens the model selector.
• As a default: the agent.model key in ~/.config/devin/config.json — %APPDATA%\devin\config.json on Windows.
• Short names resolve to the latest in a family, so opus, sonnet, swe, codex, and gemini track new releases without a config edit.
• Adaptive is a router that picks per task, and is Cognition’s recommendation for most users rather than a fixed choice.

{
  "agent": {
    "model": "swe"
  }
}

Why a compound system is built this way

It is tempting to read "does not currently support third-party LLM API keys" as a feature nobody has got around to. The phrasing suggests otherwise, and the suggestion is reasonable. A tool that is a chat loop around one model can swap the model with a base URL, because the model is the product’s only moving part. Cognition describes Devin as a compound AI system: several models and non-model components arranged into a pipeline, tuned together, with the planning, editing, and verification stages depending on the specific behaviour of the specific models behind them.

Substituting an arbitrary endpoint into that is not a configuration change but a change to what the product is. A pipeline calibrated on one set of models does not keep its properties when an unknown one is dropped into a stage, and the failures would be attributed to Devin. Whether or not you find that convincing, it is at least a coherent position, and it explains why the answer is the same on a single-tenant VPC as it is on the shared cloud.

If the requirement is your own gateway

The requirement behind the question is usually one of three things: a specific model Devin does not offer, billing that lands on an account you control, or inference that goes through a gateway you can inspect. Every one of those is met by the agent harnesses that do take a custom endpoint — and there are several good ones.

All four run against the same Continuum key, which is created in the web app under Settings, then Account, then Inference API. Keys mint on Plus and above, so on the Free tier the Create key button stays disabled until you subscribe.

Get hosted inference

One key covers Claude Code, the Codex CLI, OpenCode, and Cursor. Plus is $25/mo.

These are not equivalents of Devin and it would be silly to pretend otherwise. Devin’s pitch is an autonomous engineer you hand a ticket to; these are harnesses you drive, in a repository you have checked out, on a machine you control. What they give you in exchange is the thing this page is about: the endpoint is yours, the model list is yours, and the bill arrives on an account you chose.

Where Devin still fits

None of the above is an argument that Devin is bad, and a comparison that only lists a competitor’s constraints is not a comparison. The managed model roster is downstream of a real product decision, and that decision buys things.

• Nothing to configure. No keys, no base URLs, no config files, no per-machine setup — which matters most in exactly the organizations where those are hardest to distribute.
• One accountable vendor. When output quality drops, there is no question of whose model or whose gateway caused it. That is a genuine operational advantage, and it is precisely what a bring-your-own-endpoint feature would give away.
• Tuned as a whole. The compound pipeline is calibrated against known models. Its behaviour is a property of the system rather than of whatever you plugged in.
• Enterprise deployment options. Single-tenant VPC and customer-managed keys address data-residency and key-custody requirements, which for many organizations are the actual constraint rather than model choice.

So the honest framing is not "Devin is closed and everything else is open". It is a trade: Devin takes model choice off your plate and takes it away from you in the same motion. If the second half of that sentence is a problem for you, you have your answer — and it is the answer at the top of this page, not one that a support ticket or an enterprise contract will change.

Questions people ask

Keep reading