Claude Code Gateway

Glitch serves as the LLM gateway for Claude Code, Anthropic’s official CLI for software engineering tasks. Pointing Claude Code at your Glitch sensor gives every prompt your team sends through the CLI:

Security detection — prompt injection, jailbreaks, PII leakage, data exfiltration attempts
Audit logging — every request and response captured per Claude Code session
Centralized policy — block, log, or alert based on your team’s detection thresholds
Per-team usage tracking — token counts, latency, and detection findings per Glitch API key

Glitch transparently proxies the Anthropic Messages API at /v1/messages and forwards your existing Anthropic credentials byte-for-byte to api.anthropic.com. No code changes to Claude Code; no proxy your team has to operate beyond the Glitch sensor.

Choose your authentication path

Glitch supports both ways Claude Code authenticates with Anthropic. Pick the page that matches how your team uses Claude Code today:

Anthropic API Key

Your team uses an ANTHROPIC_API_KEY (a sk-ant-... token) to call the Claude API directly. Glitch forwards the key unchanged to api.anthropic.com.

→ Set up with API key

Claude Pro/Max Subscription

Your team uses Claude Code with personal Pro or Max subscriptions (the “Claude account with subscription” sign-in flow). Glitch forwards the OAuth bearer Claude Code obtains during sign-in, so subscription billing stays with Anthropic while Glitch handles security and audit.

→ Set up with Max subscription

The mechanics are nearly identical between the two — both modes use the same Glitch sensor, the same ANTHROPIC_BASE_URL redirect, and the same detection pipeline. The only difference is whether you set ANTHROPIC_API_KEY (API key path) or let Claude Code’s OAuth flow populate the Authorization header for you (subscription path).

At a glance

The shape of every Claude Code → Glitch → Anthropic request:

flowchart TD
    A["<b>Claude Code (claude CLI)</b><br/><br/>Authorization: Bearer &lt;api_key OR oauth_token&gt;<br/>X-Glitch-Key: glitch_sk_...<br/>X-Claude-Code-Session-Id: &lt;session_uuid&gt;<br/>anthropic-version: 2023-06-01"]
    B["<b>Glitch sensor</b><br/><br/>• Authenticates tenant via X-Glitch-Key<br/>• Runs detection pipeline (signatures + LLM-based)<br/>• Logs SecurityEvent with session id, model, tokens, blocks<br/>• Blocks at 403 with Anthropic-shaped error if policy denies"]
    C["<b>api.anthropic.com</b><br/><br/>Authorization header forwarded byte-for-byte"]

    A --> B
    B --> C

    style A fill:#1a1a2e,stroke:#00d4ff,color:#fff
    style B fill:#0d3d4d,stroke:#00d4ff,color:#fff
    style C fill:#1a1a2e,stroke:#00d4ff,color:#fff

Endpoints supported

Endpoint	Behavior
`POST /v1/messages`	Full proxy with detection, rate limiting, and audit logging
`POST /v1/messages/count_tokens`	Authenticated passthrough (no detection)

Both streaming ("stream": true) and non-streaming responses are supported.

Headers Glitch forwards

Header	Behavior
`Authorization`	Forwarded verbatim to `api.anthropic.com`. May contain an Anthropic API key (`Bearer sk-ant-...`) or an OAuth bearer from Claude Code’s subscription flow — Glitch does not inspect or rewrite it.
`x-api-key`	Forwarded verbatim if present (alternate Anthropic auth header).
`anthropic-version`	Forwarded if set; Glitch defaults to `2023-06-01` if absent.
`anthropic-beta`	Forwarded verbatim.
`X-Claude-Code-Session-Id`	Forwarded and captured into your audit log so you can group all events from the same Claude Code session.

X-Glitch-Key is consumed by the Glitch sensor for tenant authentication and is never leaked to upstream. Secret headers (Authorization, x-api-key, Cookie, etc.) are redacted to [REDACTED] in the audit log; only the request bodies, detector findings, and session metadata are persisted.

Tuned for Claude Code traffic

Glitch’s default detection policy is calibrated against real Claude Code traffic — including the system prompt and tool descriptions Claude Code sends on every request. Common false-positive patterns (path-traversal-like substrings in tool documentation, multi-turn dev conversations with file operations) are explicitly handled by the default policy so your team can use Claude Code without tuning detector thresholds first.

Next steps

Set up with an API key — three env vars, one verification curl, and you’re routing
Set up with a Max subscription — same setup minus the API key; Claude Code’s OAuth flow handles auth
API Reference — request/response shapes, error codes, rate-limit headers