An unshipped Claude Desktop setting redirects Code session traffic to a customer's own host

A client that acts on selfHostedUrl does something narrow. It matches outgoing requests against two prefixes, /v1/code/sessions and /v1/environment_providers/, and sends those to the configured HTTPS host ^[3]. Session creation and execution-environment lookup change destination. RuntimeWire's analysis also turned up authentication plumbing and a device-file WebSocket bridge, which the configuration describes as a restricted bridge to local files ^[1][4].

No shipped build acts on the value. The schema's own description says "Unreleased: only internal builds act on it" ^[5]. An administrator who writes the field into a released copy of Desktop therefore gets nothing. The setting is one line of configuration plus one build gate away, and the build gate is Anthropic's. RuntimeWire says it did not alter that gate, execute the hidden functionality, authenticate to private services or contact the placeholder endpoint, and that what it found is implementation code; it did not observe a working deployment ^[17].

The code landed inside an eight-day window: absent from RuntimeWire's September 10 comparison package, present in the September 18 build ^[6][8]. Anthropic's notes for 2.2553.1 listed session and file-rewind fixes and reported no user-facing changes for third-party deployments ^[7]. Both statements hold at once, since no released build reads the setting.

Three days before that build, Coder announced Claude Code support for Agent Relay ^[9]. Coder said it developed the integration with support from the Claude Code team and that it was in early access with selected design partners ^[10]. Its published architecture puts a runner inside each customer-controlled workspace; the runner opens an outbound connection to Anthropic's backend, and tool calls execute in the workspace ^[11]. Under that description Anthropic still operates the agent loop and handles inference, account administration and billing, with Claude Desktop or claude.ai as the developer's interface ^[12]. Desktop driving a remotely executed Code session was already public ^[13].

Coder's runner dials out to Anthropic ^[11]; the Desktop code has the client dial a host the organization names ^[3]. RuntimeWire says both paths could belong to the same deployment architecture and that the finding does not establish separate products ^[14]. Four things stay open in the package: who would operate the configured host, whether it would forward requests to Anthropic, where inference ultimately happens, and how the setting relates to Coder's integration ^[15].

Self-hosting here means the session host, and only that. RuntimeWire found no downloadable Claude model weights, no offline inference and no customer-hosted foundation model ^[16]. The schema names version 2.110.0, released September 15 ^[18], the same date Coder announced Agent Relay ^[19]; RuntimeWire says the public documentation it cites leaves the relationship open ^[15].

For this to become the audit chokepoint platform teams keep asking for, the customer has to operate the session host and keep its logs. RuntimeWire's stated stake assumes exactly that: a customer-run session host centralizing Claude Code execution, policy and audit while Desktop stays the employee interface ^[21]. On the artifact alone, who would run that host is still open. Anthropic had not responded to RuntimeWire's request for comment by publication ^[20].