Network proxy changedthird-party/claude-desktop/network-proxy
Nearest release: v2.1.315, published 3 hours after upstream edited the page. Shown because the two are within 24 hours of each other. Nothing here says the release caused the edit.
Upstream edited this page at 30 Oct 2026 17:36 UTC, give or take a minute or two: the time comes from Anthropic’s own sitemap rather than from a commit. This site recorded the change at 30 Oct 2026 17:37 UTC.
Upstream edited
Recorded here
Lines+5added
Lines−4removed
From line
6
where the diff opens
First seen
2 Sep 2026
this site's first read of the page
Recorded edits17to this page, all time
The whole hunk
from line 6, old and new numbered
/
from line 6
66
77Three parts of the product make network connections, and they do not all resolve the proxy the same way:
88
9* **The app**: the Claude Desktop window itself, including sign-in, the connection test, Web Fetch in Chat and Cowork sessions, managed MCP servers the app connects to, and plugin marketplace sync.
10* **The agent**: the Claude Code engine the app runs for every Chat, Cowork, and Code session. It sends inference requests and, in Code sessions, also makes its own web fetches, remote MCP connections, and plugin installs.
9* **The app**: the Claude Desktop window itself, including sign-in, the connection test, Web Fetch in Chat and Cowork sessions, managed MCP servers the app connects to, and sync of the plugin marketplaces you configure.
10* **The agent**: the Claude Code engine the app runs for every Chat, Cowork, and Code session. It sends inference requests and makes its own remote MCP connections, for example to [a server that a marketplace plugin or a user-added plugin bundles](https://claude.com/docs/third-party/claude-desktop/extensions.md#remote-mcp-servers-in-plugins). In Code sessions, it also makes its own web fetches and plugin installs.
1111* **Cowork's sandboxed shell**: the isolated environment where commands the agent runs in a Cowork session (`curl`, `pip`, `npm`, and so on) execute.
1212
1313With [Unified Claude](https://claude.com/docs/third-party/claude-desktop/unified-claude.md) on, a Cowork session started from the home screen runs its commands in Cowork's sandboxed shell.
from line 79
7979| App window, in-app sign-in, connection test, model list | App proxy |
8080| Inference requests from Chat, Cowork, and Code sessions | The single proxy resolved for the inference endpoint (from the pinned key or the OS), or Claude Code managed settings if deployed |
8181| Agent web fetch, remote MCP servers, and plugin installs in Code sessions | Same single proxy as inference |
82| The agent's remote MCP connections in Chat and Cowork sessions, to servers that a marketplace plugin or a user-added plugin bundles | Same single proxy as inference |
8283| Web Fetch and managed MCP servers in Chat and Cowork sessions | App proxy (the app makes these connections) |
83| Plugin marketplace sync and `aws` CLI calls the app makes for Bedrock sign-in | App proxy, resolved for the specific host |
84| Sync of the plugin marketplaces you configure, and `aws` CLI calls the app makes for Bedrock sign-in | App proxy, resolved for the specific host |
8485| Cowork sandbox download from `downloads.claude.ai` | App proxy |
8586| Telemetry and crash reports to Anthropic, if enabled | App proxy |
8687| OpenTelemetry export to your collector | App proxy for the app's own events; the same single proxy as inference for Claude Code metrics and logs |
from line 113
112113
113114### The agent uses one proxy
114115
115The agent sends all of its own traffic (inference, and in Code sessions its web fetches, remote MCP connections, and plugin installs) through the one proxy resolved for your inference endpoint. Per-host PAC rules are not re-evaluated inside the agent. For most organizations this is the desirable outcome: everything the agent does is logged and inspected at one place, with nothing further to configure.
116The agent sends all of its own traffic (inference, its remote MCP connections, and in Code sessions its web fetches and plugin installs) through the one proxy resolved for your inference endpoint. Per-host PAC rules are not re-evaluated inside the agent. For most organizations this is the desirable outcome: everything the agent does is logged and inspected at one place, with nothing further to configure.
116117
117118It needs attention only when that one proxy is not right for everything the agent reaches: typically some hosts (an internal MCP server, a private package registry, a self-hosted plugin marketplace) must be reached directly while the proxy only carries public traffic, or the reverse. In order of preference:
118119
No line in this hunk matches that.