Skip to content

docs: design for first-party Claude Desktop through agentgateway - #185

Draft
ilackarms wants to merge 2 commits into
mainfrom
docs/claude-desktop-first-party-interception
Draft

ilackarms wants to merge 2 commits into
mainfrom
docs/claude-desktop-first-party-interception

Conversation

@ilackarms

Copy link
Copy Markdown
Collaborator

A design proposal for #182: Claude Desktop stays signed in to claude.ai, and the daemon sends its Anthropic traffic through agentgateway without third-party inference mode.

The doc is docs/design/claude-desktop-first-party-interception.md. It covers:

  • the results of a macOS spike (proxy and PAC support, trust store, no certificate pinning, protocols, fail-closed behavior)
  • the intercept listener, the name-constrained device CA, and steering with a system PAC or Claude's managed key
  • forwarding to agentgateway with an identity header, and what agentgateway needs
  • a configuration sketch, security notes, the test plan, work items and open questions

This PR adds only the doc. There is no code change. Comments on the open questions in section 10 are especially welcome.

Refs #182

Propose how the daemon can send first-party Claude Desktop traffic
through agentgateway without third-party inference mode: a loopback
intercepting proxy, a name-constrained device CA, PAC steering,
forwarding with an identity header, and fail-closed behavior. The doc
also has the spike results it builds on, a configuration sketch, work
items and open questions.

Refs #182
Name the components that the spike tested, mark the Cowork VM traffic
and Chat mode as not tested, and say that the fail-closed result is for
the local proxy only, because the test had no gateway.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant