Proxy live ยท Python & TypeScript SDKs ยท 20 languages ยท 4 compliance frameworks

Your AI agents can
delete your database.

Clampd sits between your agents and the tools they use. Every database query, API call, file access, and shell command - inspected, scoped, and enforced before it executes. If an agent goes rogue, we kill it in under 25ms. One line to integrate. Self-hosted. Your data never leaves your network.

View Docs → See how it works
$ pip install clampd  /  npm install @clampd/sdk
clampd proxy - what it looks like
# Agent attempts destructive SQL via LangChain
agent โ†’ database.query("DROP TABLE users CASCADE")
 
โœ— BLOCKED - destructive SQL detected
  risk_score: high   latency: 1ms
  audit: logged
 
# Safe query passes through security pipeline
agent โ†’ database.query("SELECT name FROM users WHERE role = 'admin'")
 
โœ“ ALLOWED - clean, scope granted
  access: least-privilege scope applied
  forwarded to downstream tool   latency: 14ms
The Problem

Agents have keys to everything.

AI agents call tools with full credentials. One prompt injection turns your helpful assistant into an attacker with insider access.

๐Ÿ’‰

Prompt Injection

Malicious instructions hidden in retrieved documents redirect your agent to execute harmful tool calls.

๐Ÿ”“

Privilege Escalation

Agents inherit broad OAuth scopes. One compromised chain grants access to delete databases, send emails, transfer funds.

๐Ÿ“ค

Data Exfiltration

Injected payloads silently exfil sensitive data through tool calls - API keys, PII, credentials sent to attacker endpoints.

๐Ÿ”—

Tool Chain Attacks

Multi-step agents amplify risk. A single poisoned step cascades through the entire chain with compounding damage.

๐Ÿ‘ป

Shadow Agents

Developers spin up agents with API keys. No logging, no approval, no audit trail. Your CISO doesn't know they exist.

โฑ๏ธ

No Kill Switch

When an agent goes rogue, there's no way to cut access in real-time. By the time you notice, the damage is done.

How It Works

One security layer between your agents and their tools.

Every call is classified, evaluated, and audited - before it reaches the downstream service. Sub-10ms median latency.

01

Authenticate & Identify

Verify credentials, resolve agent profile, check kill-switch status and scoped permissions. Reject unauthenticated requests immediately.

02

Classify & Evaluate

Rules engine matches patterns against known threats. Policy engine allows or denies based on risk score, agent scope, and session context. Violations blocked in real-time.

03

Scope & Forward

Apply least-privilege access - each call is scoped to only the permissions it needs. Forward to the downstream tool and capture the response.

04

Audit

Every event recorded. Tool name, risk score, policy action, denial reason, latency - full audit trail. See full architecture →

<10ms
p95 Tool Call Latency
Multi
Stage Pipeline
18
Tool Categories Protected
2
SDKs (Python + TS)
Capabilities

Defense in depth - prompt scanning AND tool-call control.

Every prompt is scanned and every tool call is classified, scoped, and enforced. Covering all 18 tool-call categories with behavioral anomaly scoring, scope tokens, and cross-agent correlation - all sub-10ms.

Multi-Stage Security Pipeline

Every tool call passes through authentication, classification, policy enforcement, and audit. Ed25519-signed scope tokens cryptographically bind each approval to the specific operation. Under 10ms p95.

Cryptographic scope enforcement

Multi-Framework SDKs

Python + TypeScript. OpenAI, Anthropic, LangChain, Google ADK, MCP. 1 line to integrate with any agent framework.

PyPI + npm

Multilingual Threat Detection

Multi-layer detection engine covering injection attacks, data exfiltration, privilege escalation, rogue agent behavior, and encoding evasion across 20 languages. Import your own rules in Sigma YAML format. Define policies in Cedar. Custom keyword dictionaries and rule packs configurable from dashboard.

Sigma • Cedar • 20 languages

MCP Proxy Server

Wrap any MCP server - filesystem, database, GitHub, and more. Every tool call governed. Works with Claude Desktop out of the box.

MCP compatible

Emergency Kill Switch

Multi-layer cascade terminates rogue agents across all services. Tokens revoked, sessions cleared, credentials invalidated. Idempotent and audited. Typically fires in under 25ms.

<25ms termination

Full Audit Trail

Every event recorded - tool name, risk score, policy action, denial reason, latency. Query with SQL. PII automatically masked.

SQL-queryable + PII masking

Dashboard Control Plane

Kill, suspend, or pause any agent from the dashboard. Manage policies, edit rules, monitor the live risk feed with real-time updates.

Real-time monitoring

Agent Registry

Register every agent with scoped credentials and per-agent policy. No more shadow agents running untracked with shared API keys.

Scoped credentials

Prompt & Response Scanning

Full prompt scanning via /v1/scan-input runs detection rules across all 18 tool-call categories on prompts before they reach the LLM. Response inspection catches PII leaks, data exfiltration, and poisoned responses after tool execution. Both enabled by default in SDKs.

scan_input + scan_output

A2A Rug-Pull Detection

Detects when a tool's behavior changes after approval - SHA-256 descriptor hashing catches tampering immediately. Agent-to-agent delegation chains are validated with depth limits and cycle detection. Every inter-agent hop is verified and scoped.

Tool tampering detection

Behavioral Anomaly Detection

Continuous risk scoring with behavioral baselines. Detects multi-step attack chains, velocity anomalies, scope violations, and coordinated patterns. Auto-suspend on threshold breach, permanent kill on repeated violations.

Continuous risk scoring

Scope-Based Permissions

Fine-grained scope enforcement with configurable exemptions and time-limited approvals. Cryptographic scope tokens bind each approval to the specific operation. Admin configures from dashboard - developer changes nothing.

Least-privilege enforcement

Red Team Testing

Built-in CLI attack suite: clampd test. Pre-built attack vectors across multiple categories. Security teams evaluate detection coverage in minutes.

Built-in attack suite

Slack & PagerDuty Alerts

Auto-formatted alerts for Slack (Block Kit) and PagerDuty (Events API v2). Risk-based severity mapping. "Agent tried DROP TABLE - blocked" appears in your incident channel in real-time.

Auto-detect integration type

Payment Protocol Enforcement

Intercepts Google AP2 mandates and Coinbase x402 HTTP 402 responses before agent wallets transact. Per-transaction limits, hourly spend caps, vendor whitelists. Supports 8 USD stablecoins across 7 blockchain networks.

AP2 + x402

Multi-Tenant Isolation

Full organization-based multi-tenancy. Agents, API keys, policies, scope exemptions, and audit logs are all isolated per org. Built for platform teams and MSPs managing multiple product squads.

Org-based isolation

Compliance Reports

Auto-generated compliance report templates aligned to HIPAA, GDPR, SOC 2, and ISO 27001 frameworks. Control mapping with pass/fail status. Luhn credit card validation, 18 HIPAA PHI identifiers, data classification (Restricted/Confidential/Internal/Public). Evidence collection from tool call audit data.

4 compliance frameworks
Pricing

Start Free, Scale with Enterprise

All security features included on the free tier (up to 25 agents). No credit card required.

Design Partner
Design Partner
$0 - free

All features included. Same security pipeline as Enterprise with limited agent count.

  • Full 9-stage security pipeline
  • Python & TypeScript SDKs
  • 243 detection rules + Cedar custom policies
  • Emergency kill switch
  • Dashboard UI (kill / suspend / policies / live risk)
  • 7-type anomaly detection & auto-suspend
  • Prompt & response scanning (PII, secrets)
  • A2A delegation detection
  • Payment protocol enforcement (AP2 + x402)
  • PII masking & quarantine
  • Scope tokens (Ed25519)
  • Compliance reports (HIPAA, GDPR, SOC 2, ISO 27001)
  • CLI tool + red team testing
  • Webhook notifications
  • Slack & PagerDuty alerts
  • Hardened Docker images
  • Agents Up to 25
  • Requests 500,000 / month
  • API keys Up to 5
  • Support Email
Get Started Free
License-Based
Low-Latency Native Code
Docker Compose & Helm / Kubernetes
Sub-10ms Typical Latency
Get Started

Start securing your AI agents today

Sign up for free. No credit card required. Install the SDK and start guarding tool calls in minutes.

Sign Up Free Read the Docs Try the Playground

Enterprise? support@clampd.dev