anthropic-proxy

Transparent reverse proxy for Claude Code API calls to api.anthropic.com.

What It Does

anthropic-proxy sits between Claude Code (or any Anthropic API client) and api.anthropic.com, forwarding all requests over HTTPS with HTTP/2. It provides request metrics, token tracking, a live dashboard, and an optional secret scrubber that detects and pseudonymizes credentials before they reach the upstream API.

Quick Start

# Install dependencies (requires uv)
make install

# Generate a self-signed TLS certificate
openssl req -x509 -newkey rsa:2048 -keyout server.key -out server.crt \
  -days 365 -nodes -subj "/CN=localhost"

# Run locally
make run

The proxy listens on https://127.0.0.1:8443 by default. Point your client at this address instead of https://api.anthropic.com.

Configuration

All settings are controlled via environment variables prefixed with ANTHROPIC_PROXY_. Pydantic Settings handles parsing and validation.

Core

Variable	Default	Description
ANTHROPIC_PROXY_HOST	0.0.0.0	Bind address
ANTHROPIC_PROXY_PORT	8443	HTTPS listen port
ANTHROPIC_PROXY_DASHBOARD_PORT	9090	Plain HTTP port for the dashboard
ANTHROPIC_PROXY_SHUTDOWN_TIMEOUT	30.0	Seconds to wait for active requests to drain on shutdown

Logging

Variable	Default	Description
ANTHROPIC_PROXY_LOG_LEVEL	info	Python log level (debug, info, warning, error)
ANTHROPIC_PROXY_LOG_REQUESTS	false	Log every proxied request/response
ANTHROPIC_PROXY_LOG_SLOW_THRESHOLD	5.0	Seconds; requests slower than this are logged at WARNING

Upstream

Variable	Default	Description
ANTHROPIC_PROXY_UPSTREAM_BASE_URL	https://api.anthropic.com	Upstream API base URL
ANTHROPIC_PROXY_UPSTREAM_CONNECT_TIMEOUT	10.0	TCP connect timeout (seconds)
ANTHROPIC_PROXY_UPSTREAM_READ_TIMEOUT	300.0	Read timeout (seconds)
ANTHROPIC_PROXY_UPSTREAM_WRITE_TIMEOUT	30.0	Write timeout (seconds)
ANTHROPIC_PROXY_UPSTREAM_POOL_TIMEOUT	10.0	Connection pool timeout (seconds)
ANTHROPIC_PROXY_UPSTREAM_MAX_CONNECTIONS	20	Max concurrent connections to upstream
ANTHROPIC_PROXY_UPSTREAM_MAX_KEEPALIVE_CONNECTIONS	10	Max keepalive connections in pool

Secret Scrubber

Variable	Default	Description
ANTHROPIC_PROXY_SCRUB_MODE	off	off or pseudonymize
ANTHROPIC_PROXY_SCRUB_HMAC_KEY	(empty)	64-char hex key for HMAC pseudonymization
ANTHROPIC_PROXY_SCRUB_MAX_BODY_SIZE	10485760	Max request body size to scrub (bytes; 10 MB)

Secret Scrubber

The scrubber intercepts POST request bodies, detects secrets and identifiers using regex-based detectors, and replaces them with deterministic placeholders before forwarding to the upstream API.

Enabling

Set two environment variables:

ANTHROPIC_PROXY_SCRUB_MODE=pseudonymize
ANTHROPIC_PROXY_SCRUB_HMAC_KEY=<64-char hex key>

Key Generation

python3 -c "import secrets; print(secrets.token_hex(32))"

Detectors

The scrubber runs 16 detectors in priority order. Earlier detectors take precedence on overlapping matches.

#	Detector	Example Match	Tier
1	anthropic_key	sk-ant-api03-abc...	SECRET
2	openai_key	sk-proj-abc123...	SECRET
3	aws_access_key	AKIAIOSFODNN7EXAMPLE	SECRET
4	github_token	ghp_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx	SECRET
5	google_api_key	AIzaSyA1234567890abcdefghijklmnopqrstuv	SECRET
6	stripe_key	sk_live_abc123def456ghi789...	SECRET
7	private_key	-----BEGIN RSA PRIVATE KEY-----	SECRET
8	jwt	eyJhbGci...eyJzdWIi...signature	SECRET
9	db_connection_string	postgres://user:pass@host/db	SECRET
10	bearer_token	Bearer eyJhbGciOiJIUz...	SECRET
11	basic_auth	Basic dXNlcjpwYXNz...	SECRET
12	sonarqube_token	sqp_a1b2c3d4e5f6...	SECRET
13	jenkins_api_token	11abcdef0123456789abcdef01234567	SECRET
14	ansible_vault	$ANSIBLE_VAULT;1.1;AES256	SECRET
15	hashicorp_vault_token	hvs.ABCDEFghijklmnop123456	SECRET
16	email	user@example.com	IDENTIFIER

Placeholder Formats

Secrets (counter-based, per request): Replaced with <PLACEHOLDER_N> where N increments per type within a single request.

sk-ant-api03-abcdef...  ->  <ANTHROPIC_KEY_1>
ghp_xxxxxxxxxxxx...     ->  <GITHUB_TOKEN_1>

Identifiers (HMAC-based, deterministic): Replaced with <PLACEHOLDER_HMAC:HEXHEX> using a truncated HMAC-SHA256 digest. The same input always produces the same placeholder (given the same key), enabling correlation across requests.

user@example.com  ->  <EMAIL_HMAC:A3F1B2C0>

What the Scrubber Skips

• JSON skip-set keys: model, role, type, stop_sequences, temperature, top_p, max_tokens, top_k, stream, metadata, input_schema. Values under these keys are never scrubbed.
• Tool definition keys: name and description are additionally skipped inside tools arrays.
• Image data: Base64 image payloads ({"type": "base64", "data": "..."}) are left untouched.
• Non-JSON bodies: Bodies that fail JSON parsing are passed through unchanged.
• Oversized bodies: Bodies exceeding SCRUB_MAX_BODY_SIZE (default 10 MB) are passed through unchanged.
• Non-POST requests: Only POST request bodies are scrubbed.

Endpoints

Endpoint	Port	Description
GET /health	8443	Returns {"status": "ok"}
GET /metrics	8443	JSON snapshot of request counts, error rates, bytes, tokens, scrub stats
GET /metrics/stream	8443, 9090	SSE stream of metrics snapshots (every 1.5s)
GET /dashboard	8443, 9090	Live HTML dashboard
* /{path}	8443	Catch-all reverse proxy to upstream

Dashboard

The live dashboard is a single-page HTML app served at /dashboard on both the main HTTPS port and the plain HTTP dashboard port (default 9090). It connects via SSE to /metrics/stream and displays:

• Requests per second and output tokens per second (time-series charts, last 60 samples)
• Error rate, active requests, uptime
• Data throughput (bytes in/out) and token totals (input/output)
• Status code breakdown table
• Alert banner for high error rate (>10%) or high active requests (>10)

Deployment

An Ansible role is included under deploy/ansible/. It handles:

1. Creating a dedicated system user (anthropic-proxy)
2. Syncing source code to /opt/anthropic-proxy/app/
3. Installing uv and running uv sync
4. Deploying a .env file and systemd service unit
5. Configuring UFW firewall rules for both ports
6. Health-checking the service after startup

# Deploy to hosts defined in deploy/ansible/inventory.ini
make deploy

Development

Make Targets

Target	Description
make install	Install dependencies with uv sync
make run	Run proxy locally with debug logging and self-signed TLS
make deploy	Deploy via Ansible
make logs	Tail remote service logs
make status	Show remote systemd service status
make health	Curl the /health endpoint
make metrics	Curl the /metrics endpoint

Running Tests

uv run pytest

Project Structure

src/anthropic_proxy/
    __init__.py
    config.py          # Settings and validation (pydantic-settings)
    main.py            # FastAPI app, proxy logic, metrics, dashboard server
    scrubber.py        # Secret detection and pseudonymization
    dashboard_html.py  # Inline HTML/JS/CSS for the live dashboard
tests/
    test_scrubber.py
deploy/ansible/
    playbook.yml
    inventory.ini
    roles/anthropic_proxy/
        tasks/main.yml
        templates/      # .env and systemd unit templates
        defaults/main.yml
        handlers/main.yml

Architecture

1. A client sends a request to anthropic-proxy over HTTPS.
2. The proxy reads the request body. If scrubbing is enabled and the request is a POST, the scrubber parses the JSON body, walks the structure (skipping known-safe keys and image data), and replaces detected secrets with deterministic placeholders.
3. The proxy forwards the request to api.anthropic.com via an HTTP/2 connection pool, streaming the response back to the client.
4. On 529 (overloaded) responses, the proxy retries up to 3 times with exponential backoff (1s, 2s, 4s).
5. Metrics (request counts, status codes, bytes, tokens, scrub hits) are recorded and exposed via /metrics and the SSE stream powering the dashboard.