Skip to main content

How to govern ChatGPT/Claude Desktop (enterprise)

This guide describes how enterprise IT can route ChatGPT Desktop and Claude Desktop traffic through Talon. Every conversation from managed devices is then audited and policy-enforced.

Prerequisites: Enterprise-managed devices, Talon deployed on an internal server, and (for the interception phase) an internal CA plus DNS control.

How it works​

  1. Desktop clients (or the enterprise network path) send OpenAI/Anthropic-bound traffic to the Talon gateway.
  2. Talon resolves the presented agent key to an agent identity — one agent.talon.yaml per desktop fleet or team, bound to a vault key (#266).
  3. The gateway pipeline runs (PII scan, effective policy, evidence); Talon forwards to the real provider using vault-stored provider keys.
  4. Employees use ChatGPT Desktop or Claude Desktop as usual.

Architecture​

[Employee laptop] → [Talon gateway]
                     → agent key → agent identity (per team/fleet)
                     → PII scan, effective policy, signed evidence
                     → Forward to real api.openai.com with vault key

Identity for desktop fleets is the same as for every other workload: a Talon agent key. There is no source-IP identity — identify_by: source_ip and trusted_proxy_cidrs were removed in the #266 cutover (network context is not use-case identity; a spoofable address must never select policy). Network-bound identity, if it returns, will be mTLS — an explicit future feature.

Configuration​

One agent file per desktop fleet or team:

# chatgpt-desktop-engineering.talon.yaml
agent:
  name: chatgpt-desktop-engineering
  version: "1.0.0"
  tenant_id: default
  key:
    secret_name: chatgpt-desktop-engineering-talon-key

policies:
  cost_limits:
    daily: 15.00

metadata:
  team: engineering

Bind the traffic key and start the gateway:

talon secrets set chatgpt-desktop-engineering-talon-key <generated key>
talon serve --gateway

Distribute the agent key to the managed devices via MDM/GPO as the API key the desktop client presents (for clients that support a custom base URL + key), pointing them at http://talon:8080/v1/proxy/openai/v1.

The gateway config also carries a network_interception schema block for the DNS/TLS interception phase (not yet implemented):

gateway:
  # ...
  network_interception:
    enabled: false   # Phase 2
    intercept_hosts:
      - original: "api.openai.com"
        provider: "openai"
      - original: "api.anthropic.com"
        provider: "anthropic"
    tls:
      cert_dir: "/etc/talon/certs/"

Steps (when network interception is implemented)​

  1. Deploy Talon on an internal server.
  2. Generate TLS certificates for api.openai.com and api.anthropic.com using your internal CA.
  3. Deploy the internal CA to managed devices (e.g. via MDM/GPO).
  4. Update internal DNS so api.openai.com and api.anthropic.com resolve to the Talon server.
  5. Store real provider API keys in Talon's vault.
  6. Configure the gateway with network_interception enabled and one agent identity per fleet.
  7. Verify: an employee uses ChatGPT Desktop or Claude Desktop; Talon logs every request and applies policy.

Note that interception only moves the transport into Talon; identity remains the agent key. Desktop apps that cannot present a key cannot be attributed per-team until the mTLS identity feature exists — govern them behind a per-team egress proxy that injects the team's agent key.

Compliance outcome​

  • Every ChatGPT/Claude conversation from managed devices is audited.
  • PII in prompts is detected and logged (or blocked/warned per policy).
  • Cost is tracked per team via the agent identity's metadata.team.
  • Policy violations can be blocked or limited by model and cost.

Current status​

  • Now: Use the gateway as a proxy by pointing clients at http://talon:8080/v1/proxy/openai/v1 (trailing /v1 so paths like chat/completions become .../v1/chat/completions) with the fleet's agent key.
  • Phase 2: Full DNS interception (Talon as TLS endpoint for api.openai.com / api.anthropic.com) so desktop apps need no config change.

You're done​

You now know how to route ChatGPT/Claude Desktop traffic through Talon (proxy with an agent key today; DNS interception later). Talon logs every request and applies the fleet agent's effective policy.

Next steps:

I want to…Doc
Cap cost per team or applicationHow to cap daily spend per team or application
Route a Slack bot or script through TalonAdd Talon to your existing app, Slack bot
Export evidence for auditorsHow to export evidence for auditors
Understand the gateway pipelineWhat Talon does to your request