99.8% detection on a public, reproducible benchmark

Runtime Security for AI Agents: Transparent Local Proxy

Crawdad sits between your agent and its model as a transparent local proxy. Multi-layer detection pipeline inspects every request. Raw content never leaves your machine by default. Works with any provider, any framework.

terminal
$ curl -fsSL https://getcrawdad.dev/install.sh | sh
Crawdad Dashboard — real-time detection overview, threat categories, agent activity

Running in seconds.

Install, point your agent's base URL at the local proxy, activate your license. Crawdad intercepts every API call transparently — your agent doesn't know it's there.

  • One install command — macOS, Linux, WSL
  • Route via ANTHROPIC_BASE_URL or equivalent
  • Cooperative path — works alongside your existing workflow
  • No code changes to your agent
quickstart
# 1. Install
$ curl -fsSL https://getcrawdad.dev/install.sh | sh

# 2. Route your agent through Crawdad
$ export ANTHROPIC_BASE_URL=http://localhost:8080/anthropic/v1

# 3. Activate
$ crawdad setup --license

# That's it. Every request is now inspected.

Multi-layer inspection. Every request.

Not a single classifier. Not a regex list. A multi-layer prompt injection protection pipeline where each layer catches what the others miss. Built in Rust. 2,988 tests across 26 crates.

Request In agent → proxy
L1 Patterns 127 / 22 categories
Behavioral 7 sub-checks
DeBERTa 44M classifier
Arbiter final verdict
Model or blocked
L1

Pattern Matching

127 patterns across 22 attack categories. Catches known injection formats, encoding tricks, and role-override attempts before anything else runs.

127 patterns · 22 categories
Behavioral

Behavioral Analysis

7 sub-checks analyze structural signals — conversation topology, authority escalation, instruction-data boundary violations. Catches attacks that look benign to static patterns.

7 sub-checks
Indirect

Indirect Injection

Detects instructions hidden inside documents, tool outputs, and retrieved content. The threat that enters through what your agent reads, not what the user types.

Session

Session Context

Maintains a 20-message sliding window to catch multi-turn attacks that unfold gradually — slow escalation, context poisoning, progressive boundary shifting.

20-message window
PII / Cred

Credential & PII Detection

Scans for 10 credential and PII types — API keys, tokens, SSNs, credit cards — before they reach the model. Blocks exfiltration at the proxy layer.

10 credential types
ML

DeBERTa-v2-small Classifier

44M-parameter transformer fine-tuned on adversarial prompt data. Runs locally on-device. Catches novel attacks that evade pattern and behavioral layers.

DeBERTa-v2-small · 44M params

The Arbiter combines signals from every layer into a single verdict. Unanimous agreement — not majority vote — determines final disposition.

GET /v1/detections
POST /v1/analyze
GET /v1/sessions/{id}/events
POST /v1/policies
GET /v1/benchmark/results
DEL /v1/sessions/{id}

Full reference in Redoc →

Full REST API. Documented in Redoc.

Detections, sessions, policies, audit logs, fleet management — every capability exposed through a documented REST API with OpenAPI 3.0 spec and interactive Redoc reference.

  • OpenAPI 3.0 specification
  • Interactive Redoc documentation
  • Detections, sessions, policies, audit
  • Fleet management endpoints
API Reference → Contact Us for API Access

Works with what you already use.

Crawdad is a transparent proxy — an LLM firewall that sits between your agent and the model. If your tool talks to an LLM API, Crawdad can inspect it. No SDK, no code changes.

Agents & Tools
Claude Code Anthropic CLI
Windsurf AI IDE
Cline VS Code Extension
Aider Terminal Agent
Continue IDE Extension
Providers
Anthropic Claude
OpenAI GPT / o-series
Google Gemini
xAI Grok
NVIDIA NIM
99.8% detection rate on a public, reproducible corpus
497attacks tested
1,669total samples
0.09%false-positive rate
CC BY 4.0corpus license
Every result published. Clone the corpus, run the benchmark, compare your tool. Every number is reproducible.

497 attack samples, 1,172 benign negatives, 22 categories. The entire corpus is open under CC BY 4.0. The product is licensed under BSL 1.1.

AndrewSispoidis/contemporary-agent-attacks →

Local-first. Zero-knowledge by default.

Crawdad runs as a transparent proxy on your machine. Raw content — prompts, responses, tool outputs — never leaves your device by default. Metadata-only telemetry by default, customer-governed.

Your machine — raw content stays local
🤖
Your Agent
any framework
Crawdad Proxy
inspects every request
🧠
AI Provider
any model
✓ on-device detection ✓ raw content stays local ✓ transparent to your agent ✓ one env var to enable

One environment variable routes your agent's traffic through the local proxy. Your agent doesn't need to know Crawdad exists. Detection runs entirely on-device.

Ship agents your users can actually trust.

Start with the install. Read the docs. Run the benchmark yourself.

Get in Touch

Tell us about your needs and we'll follow up.