The consumer Claude desktop app has no base-URL setting and ignores the Claude Code environment variables. The supported path is Claude Desktop on third-party inference, which is the same application plus a managed configuration that activates third-party mode, with no separate download. The key comes from the Continuum web app under Settings, then Account, then Inference API, created under Your keys, which needs Plus or higher. Install the app, do not sign in, enable Developer Mode under Help then Troubleshooting, open Developer then Configure Third-Party Inference, and set Inference provider to Gateway, Gateway base URL to https://continuumcode.ai (the bare origin, because the app calls POST /v1/messages), Gateway API key to a cont_sk_ key, Credential kind to Static API key, and Gateway auth scheme to Bearer. Click Apply locally; the app relaunches and the sign-in screen offers to start in third-party mode. Anthropic frames this mode for organizations whose requirements prevent first-party inference and documents it as MDM-delivered, but the single-machine path is documented too. Misspelled configuration keys are silently ignored, so verify with Help then Troubleshooting then Copy Managed Configuration Report.
- The desktop app ignores
ANTHROPIC_BASE_URL. Environment variables are a Claude Code mechanism, not a desktop-app one. - Third-party inference mode is not a separate download: it is the same app plus a configuration that switches it on.
- The gateway base URL is the bare origin,
https://continuumcode.ai— the app callsPOST /v1/messages, which Continuum serves natively. - It is an organization-oriented mode that Anthropic documents as MDM-delivered, with a documented single-machine path for evaluation. Both halves of that are true.
- The model picker auto-discovers from
GET /v1/modelsbut shows only recognizably-Claude IDs; setinferenceModelsto expose anything else. - A misspelled key is silently ignored, not reported. The Managed Configuration Report is the only honest way to see what the app actually read.
Why the usual method does not work here
The Claude desktop app and the Claude Code CLI are different programs that happen to share a brand. Claude Code is a Messages API client whose endpoint is chosen by environment variables, which is why pointing it at Continuum takes four exports. The consumer desktop app has no equivalent: no base-URL field in Settings, no configuration file that accepts one, and no effect from setting ANTHROPIC_BASE_URL before launching it.
That mechanism is Claude Desktop on third-party inference, usually written "on 3P". Anthropic describes it as "the standard Claude Desktop application plus a managed configuration that activates third-party inference mode" — so there is nothing extra to download, and no build to obtain. Turning it on is a matter of getting configuration to the app. One of its supported inference providers is a generic gateway that implements the Anthropic Messages API, and that is what Continuum is.
What changes in third-party mode, from Anthropic’s architecture table, read 13 August 2026.
| Component | Standard desktop app | On third-party inference |
|---|---|---|
| Model inference | The Anthropic API | Your configured provider endpoint |
| Web application | Loaded from claude.ai | Bundled inside the desktop app |
| User identity | An Anthropic account | Local device identity only |
| Conversation storage | Anthropic’s backend | Local disk on your machine |
| Configuration | The admin console at claude.ai | OS-native configuration, managed or per-user |
What you are signing up for
This is the section to read before the steps, because the mode is a bigger change than a base-URL swap and it is worth knowing what you are trading.
- It needs Continuum Plus or higher. Inference keys do not mint on the Free tier, so the Create key button stays disabled until you subscribe, and the gateway fields below have nothing to hold without one.
- It is built for organizations, and reachable by individuals. Anthropic designs it for organizations whose security, regulatory, or contractual requirements prevent first-party inference, and its documentation is written around administrators supplying values through MDM. It also documents a single-machine path, explicitly for evaluation, pilots, and organizations that do not use MDM. Both of those are true at once, and this guide uses the second.
- You do not sign in. Identity is local to the device and conversations are stored on local disk. Anything that exists because of an Anthropic account — history that follows you to another device, work continuing on Anthropic’s infrastructure while your laptop is shut, picking a session back up from your phone — is not part of this deployment, because there is no account for it to hang off.
- MCP tool search is off. Third-party mode disables it by default, along with Claude Code’s other experimental beta features, because strict gateways reject the experimental request headers they add. That suppression takes precedence over the
ENABLE_TOOL_SEARCHenvironment variable, so setting that variable does nothing here. - A managed profile wins. If your machine already has an MDM profile for Claude, the configuration window opens read-only and shows the deployed values. Author on a machine without one.
Configure it on one machine
Create a Continuum API key
In the Continuum web app at continuumcode.ai/app, open Settings, then Account, then Inference API. Create your key under Your keys. The same panel is on the Usage tab. Copy it when it appears: it starts with cont_sk_ and is shown exactly once.
Keys mint on Plus and above, so on the Free tier Create key stays disabled until you subscribe.
Install Claude Desktop, and do not sign in
Download it from claude.com/download and launch it. Do not sign in or create an Anthropic account. Signing in starts the standard app; you want the sign-in screen to stay where it is while you configure the other mode.
Turn on Developer Mode and open the configuration window
From the macOS menu bar — or on Windows, the ☰ application menu at the top-left of the sign-in screen — go to Help → Troubleshooting → Enable Developer Mode, then Developer → Configure Third-Party Inference…. The window that opens validates fields as you type and shows which ones your chosen provider needs.
Fill in the Connection section
Five values. The UI name is on the left; the underlying configuration key is what you would see in an exported profile or a Managed Configuration Report.
| Field in the UI | Configuration key | Value |
|---|---|---|
| Inference provider | inferenceProvider | Gateway (gateway) |
| Gateway base URL | inferenceGatewayBaseUrl | https://continuumcode.ai |
| Gateway API key | inferenceGatewayApiKey | Your cont_sk_ key |
| Credential kind | inferenceCredentialKind | Static API key (static) |
| Gateway auth scheme | inferenceGatewayAuthScheme | Bearer (the default) |
The base URL is the bare origin with no /v1, for the same reason it is in Claude Code: the app builds POST /v1/messages itself. Continuum serves that route natively, so no translating proxy is involved. Auth scheme can be either value — the gateway accepts a bearer token and an x-api-key header — so leave it on the default.
Apply locally, and start in third-party mode
Click Apply locally. The configuration is written to the application’s local config for this device and user account, and the app relaunches. The sign-in screen now offers the option to start in third-party mode using what you entered, alongside the normal Anthropic sign-in. Choose the former to work on Continuum; choose the latter at any time to go back to the standard app.
The model picker
When inferenceModels is unset, the app populates its picker from the gateway’s GET /v1/models response. Continuum serves that endpoint, so discovery works with no extra configuration — but the app filters it. Auto-discovery shows only models whose IDs are recognizably Claude, so the claude-* family appears and the rest of the served list does not.
What auto-discovery surfaces on the paid lane, read 17 August 2026.
| Model ID | Appears automatically |
|---|---|
claude-opus-5 | Yes — the flagship |
claude-fable-5 | Yes |
claude-sonnet-5 | Yes |
gpt-5.6-sol, grok-4.6, gemini-3.5-flash, and the rest | No — list them in inferenceModels to expose them |
curl -s https://continuumcode.ai/v1/models \
-H "Authorization: Bearer cont_sk_..."When it does not work
The distinctive failure mode of this mode is silence. The app reads configuration keys by name and silently ignores a misspelled one rather than reporting an error, so a typo produces an app that behaves as though you configured nothing at all. Diagnose from the report, not from the window.
| Symptom | Cause | Fix |
|---|---|---|
| The normal claude.ai sign-in screen, no third-party option | The configuration was not read at all | inferenceProvider missing or misspelled; or the app was running when you applied — quit it fully and relaunch |
| Nothing you type seems to stick | A managed MDM profile is present, so the window is read-only | Author on a machine without a managed profile |
| The model picker is empty | Discovery failed and inferenceModels is empty | Check the key with the curl above; the picker falls back to that list, and an empty list means an empty picker |
| The picker shows fewer models than the gateway serves | The recognizably-Claude filter | Expected. Set inferenceModels to list the others explicitly |
| Everything looks right and it still will not start in 3P | A required key for the chosen provider is missing | Read the Managed Configuration Report, then the application log |
- Logs.
~/Library/Logs/Claude-3p/main.logon macOS,%LOCALAPPDATA%\Claude-3p\Logs\main.logon Windows. - On macOS, an editable configuration window means no recognized key reached the app — even if your MDM insists the profile was delivered.
- On Windows, machine policy is absolute. Any value directly under
HKLM\SOFTWARE\Policies\Claude— even a misspelled one — makes the app ignore user policy entirely. The Managed Configuration Report shows which source it actually read. - A full diagnostic bundle is under Help → Troubleshooting → Generate Diagnostic Report. It carries configuration state, logs, and environment details, and no conversation content.
Questions people ask
No. Those environment variables are a Claude Code mechanism. The desktop app has no base-URL setting and ignores them; third-party inference mode is the supported route to a different endpoint.
No. Anthropic describes it as the standard Claude Desktop application plus a managed configuration that activates third-party inference mode. You install the normal app from claude.com/download and configure it.
No. Use the bare origin https://continuumcode.ai. The app calls POST /v1/messages itself, and Continuum serves the Anthropic Messages API natively, so no translation layer is needed.
It is designed for organizations whose security, regulatory, or contractual requirements prevent first-party inference, and Anthropic documents it as MDM-delivered. It also documents a single-machine path for evaluation and for organizations that do not use MDM, which is what these steps use.
Auto-discovery shows only models whose IDs are recognizably Claude, so the claude-* family appears and the rest of the served list does not. Set inferenceModels explicitly with full model IDs to expose others; that list replaces discovery rather than adding to it.
Check the spelling of the keys. The app reads managed keys by name and silently ignores a misspelled key rather than reporting an error, so a typo looks identical to no configuration. Help then Troubleshooting then Copy Managed Configuration Report shows what was actually read.
Ask your IT department first. If a managed profile is present the configuration window opens read-only, and a policy pushed later overrides local values. On Windows, any value under HKLM\SOFTWARE\Policies\Claude makes the app ignore user policy entirely.
~/Library/Logs/Claude-3p/main.log on macOS and %LOCALAPPDATA%\Claude-3p\Logs\main.log on Windows. A fuller bundle comes from Help then Troubleshooting then Generate Diagnostic Report, which excludes conversation content.
Choose the Anthropic sign-in option on the sign-in screen instead of the third-party one. The configuration stays on disk and can be edited or removed from the same window at any time.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- Claude Desktop on 3P: overview who the mode is for, and the architecture table quoted above
- Claude Desktop on 3P: installation and setup the "standard application plus a managed configuration" definition, the single-machine steps, the silently-ignored-key behaviour, and the log paths
- Claude Desktop on 3P: in-app configuration Developer Mode, Apply locally, and the read-only window under a managed profile
- Claude Desktop on 3P: LLM gateway the gateway keys, the Messages API requirement, model auto-discovery and its Claude-only filter, and the tool-search suppression
- Continuum pricing plan names, prices, and which tiers include personal API keys
- Continuum hosted inference the model lanes the picker discovers, read 13 August 2026; the live /v1/models endpoint itself needs a cont_sk_ bearer