Skip to main content
To use Claude Cowork with abliteration.ai, configure Claude Desktop’s third-party inference mode with inferenceProvider: gateway and inferenceGatewayBaseUrl: https://api.abliteration.ai. Once configured, every Cowork and Claude Code Desktop session routes through abliteration; Anthropic never sees prompts or completions. The same Claude Desktop binary runs both standard and 3P modes — the managed configuration profile activates 3P.

​
Requirements

OSmacOS 13 (Ventura) or later, Windows 10 or 11 (Virtual Machine Platform feature enabled)
AppClaude Desktop, downloaded from claude.com/download
Gateway endpointhttps://api.abliteration.ai
API keyAn ak_... key from the console

​
Configuration keys

Cowork reads its configuration once at launch. Quit and relaunch after any change.
KeyRequiredValue
inferenceProviderYesgateway
inferenceGatewayBaseUrlYeshttps://api.abliteration.ai
inferenceGatewayApiKeyYesak_YOUR_API_KEY
inferenceGatewayAuthSchemeNobearer (default) or x-api-key
inferenceGatewayHeadersNoJSON-string array of extra "Name: Value" headers (e.g. ["X-Org-Id: team1"])
inferenceModelsNoJSON-string array. Omit to auto-discover via GET /v1/models.

​
Setup via the in-app UI

For a single device or to generate a profile for fleet rollout:
  1. Open Claude Desktop. Help → Troubleshooting → Enable Developer Mode.
  2. Developer → Configure third-party inference.
  3. Select Gateway, enter the values from the table above, save.
  4. Use Export to write .mobileconfig (macOS) or .reg (Windows) for distribution.

​
Setup via MDM

Push the same keys directly to the OS preference store.

​
macOS

Domain: com.anthropic.claudefordesktop. Deliver via Jamf, Kandji, Mosyle, or any MDM that supports App Configuration. 
<key>inferenceProvider</key>
<string>gateway</string>
<key>inferenceGatewayBaseUrl</key>
<string>https://api.abliteration.ai</string>
<key>inferenceGatewayApiKey</key>
<string>ak_YOUR_API_KEY</string>
Array-typed keys must be a <string> containing a JSON array, not a native plist <array>: 
<key>inferenceModels</key>
<string>["abliterated-model"]</string>

​
Windows

Registry path: HKCU\SOFTWARE\Policies\Claude (per-user) or HKLM\SOFTWARE\Policies\Claude (machine-wide). Deliver via Group Policy, Intune, or any MDM that supports .reg files. 
[HKEY_CURRENT_USER\SOFTWARE\Policies\Claude]
"inferenceProvider"="gateway"
"inferenceGatewayBaseUrl"="https://api.abliteration.ai"
"inferenceGatewayApiKey"="ak_YOUR_API_KEY"

​
Authentication

SchemeHeader sent
bearer (default)Authorization: Bearer ak_...
x-api-keyx-api-key: ak_...
Both are accepted on /v1/messages. For environments that disallow static keys, point inferenceCredentialHelper at an executable that prints a short-lived credential to stdout (TTL via inferenceCredentialHelperTtlSec, default 3600 s).

​
Models

When inferenceModels is unset, Cowork populates the picker from GET /v1/models. Today that surfaces abliterated-model. To pin explicitly: 
"inferenceModels": ["abliterated-model"]
The first entry is the default.

​
Verifying the install

Launch Claude Desktop on a test machine. You should see:
  • Cowork and Code tabs in the left navigation
  • No Chat tab
  • A Gateway sign-in option
If users see an error at launch, check that inferenceProvider is set and that the gateway key is valid. macOS console logs and Windows Event Viewer surface deeper errors.

​
Per-user policy

Cowork’s blanket usage caps (inferenceMaxTokensPerWindow, inferenceTokenWindowHours) apply per device. For per-user or per-team enforcement, issue a project-scoped key per user — policy rules attached to that key apply to every Cowork turn. Streaming policy metadata is delivered on every SSE frame; see streaming policy metadata.

​
Data handling

Inference traffic terminates at abliteration.ai — Anthropic never sees prompts, completions, or tool inputs. Telemetry to Anthropic is metadata only (token counts, error diagnostics) and can be disabled with disableEssentialTelemetry and disableNonessentialTelemetry. See security.

​
Reference

Last modified on May 3, 2026