Claude Code through PromptGate

This recipe routes Claude Code’s traffic through your PromptGate instance. The result: Claude Code keeps doing exactly what it does, but every prompt + response is recorded, every cost is in your dashboard, every secret leak is caught.

Prerequisites

A PromptGate instance reachable from your laptop (e.g. http://localhost:8080 for self-host or your deployed URL).
An Anthropic credential configured under Credentials.

Steps

1. Create an Agent Proxy project

Projects → Create project → name it (e.g. Engineering Egress), pick Agent Proxy as the type. Default-secure guardrails (Reversible Redaction, Secret Scanner) become available under the Guardrails tab.

2. Enable the Anthropic provider in the wrapper

Under the project’s Providers sidebar entry, toggle Anthropic on and pick the credential you set up.

3. Issue an API token

API Tokens → New token with the chat scope. Copy the plaintext (pg_live_…) — it’s only shown once.

4. Point Claude Code at PromptGate

In whatever shell you launch Claude Code from:

export ANTHROPIC_BASE_URL="https://your-promptgate-host/api/<project-uuid>"
export ANTHROPIC_AUTH_TOKEN="pg_live_..."

Restart Claude Code. From its perspective nothing changed — it’s still calling the Anthropic Messages API at /v1/messages.

5. Verify

Run any Claude Code session, then open Live Logs in PromptGate. Each request appears immediately with prompt + response + token count.

Cost optimisation

The killer follow-up: route small prompts to a cheaper model. Define a wrapper alias claude-fast that points at anthropic:claude-haiku-… and have Claude Code use the fast model for trivial tasks:

# In the same project, add a routing rule on the alias:
#   when input_tokens < 4000  → claude-haiku
#   else                       → claude-3-5-sonnet

Per-team experience varies, but typical savings on coding-agent traffic are 70-85% because the bulk of requests are small “what does this function do?” prompts that Haiku handles fine.

Security follow-up

Enable Reversible Redaction with email, phone, and a few custom_patterns for your customer-id format. Your developers paste real customer data into Claude Code; the LLM only ever sees [[CUSTOMER_001]].
Enable Secret Scanner in block mode. The first time someone copy-pastes a .env file with an AWS key into Claude Code, PromptGate intercepts loudly.
Subscribe to webhooks for endpoint.anomaly so a sudden spend spike pages whoever is on call.

Troubleshooting

401 from Claude Code: the ANTHROPIC_AUTH_TOKEN env doesn’t carry a valid pg_live_… token, or the token’s scopes don’t include chat.
400 “This project is not an AI Wrapper or Agent Proxy”: the token was issued under an ai_gateway project. Tokens are project-scoped — issue a new one under the Agent Proxy project.
All requests blocked: check Guardrails → Secret Scanner. Block mode is loud by design; switch to redact if you want to keep working with secret-bearing prompts.

© Akyros Labs LLC. All rights reserved.