>50%>012345678901234567890123456789012345678901234567890123456789% of all AI usage is code. We manage that line.

Nothing today. Selan is in early access. We charge nothing and invoice nothing, and if that ever changes you'll be told well in advance and asked to agree before any charge. What is never free is your provider usage: Anthropic, OpenRouter and anyone else you connect bill you directly, on your own account, at their prices. We never see that money.

No. selan claude instead of claude; every flag and argument passes through unchanged. Claude Desktop and Codex run on your company's accounts the same way.

Yes. Models your company enables appear in the /model picker, and any model your provider serves can be named with --model provider:model-id.

Yes. Your CLAUDE.md, AGENTS.md and skills keep working: Selan sits below the harness, not inside your prompts. The one thing it does touch is MCP. Connectors your company switches on are added to your ~/.claude.json, and removed again when the company switches them off. Your own entries are left alone.

There is some added latency. Selan buffers the full request body before forwarding it, and with secret scanning on it holds the body long enough to sweep it. Nothing on the way back is scanned.

If our sign-in service is unreachable, selan claude warns and launches anyway on your cached session. The proxy is different: Selan sits on the request path, so if it is unavailable, requests through it fail. Your providers can still be reached directly with your own credentials, which is one reason the credential stays yours.

With secret scanning on, the request body is buffered at the proxy and swept with Gitleaks' own rule set (219 rules, vendored instead of fetched over the network) before anything goes upstream. A match is replaced with a placeholder in flight, so the provider never sees the real value. It redacts and forwards; it never blocks. Scanning is off by default and turned on per company.

Selan is a proxy, so it receives the full request body: your prompts and the files Claude Code attaches. None of it is logged or stored; what we keep is metadata about the call. One exception, and it is ours to name: when a provider rejects a request, up to 2 KB of its reply is written to our request log.

In the EU. The proxy runs in Warsaw (europe-central2). Request bodies are not stored; the usage ledger holds metadata, not bodies.

A revocable Selan session token, and no provider key of any kind: no Anthropic key, no Vertex or OpenRouter credential, no OAuth account. Revoke it and the machine is out. Around it sit the CLI's own files: an update check, a usage cache, a settings file that carries the same session token so it never rides on a command line, and, if you use them, a block in Claude Code's MCP config and one in Codex's config. The full list is in the reference.

The proxy governs what goes through it. For fleet-level enforcement we compose with Anthropic's managed settings so policy is applied at the OS level, not just the CLI.

The accounts are yours. Selan leases credentials you added and never owns them, and providers bill you directly on your own account. Developers drop the selan prefix and they are back on claude directly.