Agentic Runtime Security

AI Agents act on your behalf.
We make sure they don't go rogue.

One line of code. Full visibility, spend control, and compliance paperwork — before your agent acts.

👁
See everything your AI does.
🔔
Get a Slack alert before it spends money.
🛡
Automatic legal protection.
🔧
Fix security holes instantly.
Live Protection Feed
BLOCKED exec_sql — DROP TABLE detected ~$12 saved
LOOP http_request — Repeated 8x in 15s ~$34 saved
BLOCKED read_env — Credential access attempt ~$0 saved
LOOP send_email — Repeated 5x in 10s ~$8 saved

Get Started in 60 Seconds

Fork it. Run it. Done.

Full governance engine, 22 security rules, real-time Kill Feed — running on your own infrastructure in one click.

Fastest
🏠

Self-Hosted

Fork on Replit and start protecting your agents in under 60 seconds. Full source code, your infrastructure, your data.

  • One-click fork — runs instantly
  • 6 rules pre-loaded, 30+ available
  • Full source code (Apache 2.0)
  • Your data never leaves your server
Fork on Replit — Free →

Or deploy with Docker

git clone https://github.com/snapwire-ai/snapwire.git && cd snapwire && docker compose up
Coming Soon

Cloud (Coming Soon)

Fully managed. We'll handle infrastructure, updates, and scaling. Join the waitlist to get early access.

  • Instant setup, no servers needed
  • Automatic updates & scaling
  • Free — Bring Your Own Key
  • Dashboard & analytics included
Join Waitlist →

What to Expect

From fork to protected in 4 steps

Here's exactly what happens after you click "Fork." No surprises, no hidden setup.

1

Fork

One click on Replit and your own Snapwire instance spins up. Full source code, your infrastructure, running in seconds.

2

Setup Wizard

Set your admin email, choose your first security rules, and configure budget limits. Guided walkthrough — no docs required.

3

Send Your First Tool Call

Copy-paste a curl command, hit enter, and watch your first Snap-Card appear in the review queue. Instant proof it works.

curl -X POST /api/intercept -d '{"tool_name":"send_email"}'
4

You're Protected

Your dashboard lights up with real-time monitoring. All deterministic features — loop detection, spend tracking, kill switch — are active immediately.

Fork on Replit — Free →

How It Works

60 seconds to your first safety net

One line of code. Total visibility. Any agent framework.

1

Plug In

Add our proxy URL to your agent's config. One environment variable and you're connected. Works with any LLM provider. Compatible with LangChain, CrewAI, OpenAI Assistants, and any MCP-compatible agent.

OPENAI_BASE_URL=https://your-snapwire/v1
LangChain CrewAI OpenAI Assistants MCP A2A
2

Watch

Every tool call is intercepted and scored for risk. Snapwire checks for hallucination loops, spend overruns, data exfiltration, schema violations, and goal drift — all before the action executes.

Fuse Breaker — kills runaway loops
💰 Burn Meter — tracks spend in real time
🔍 Taint Tracking — follows sensitive data flow
3

Control

High-risk actions are held for human review as Snap-Cards. One-click approve, deny, or edit-and-release. Automatic NIST IR 8596 compliance reports. Fix prompts generated for your agent. Every decision logged for audit.

📋 Snap-Cards — full-context review queue
🔧 Fix Prompts — AI-generated corrections
📄 NIST compliance & audit trail

Core Features

Built-in safeguards. Every call. Every agent. Day one.

No configuration required. These work out of the box, with any LLM, on any framework.

💰

Live Burn Meter

Know exactly what your agent is spending, right now. Set daily and per-session limits. Get alerts before a runaway loop drains your credits — not after.

🔄

Loop Breaker

Agent repeats the same call 3 times in 30 seconds? Circuit broken, instantly. No more waking up to a $200 bill because your agent got stuck in a "thinking" loop overnight.

🔐

Credential Proxy

Your agent never touches a raw API key. Snapwire proxies credentials through revocable tokens. Revoke one tool, or hit the kill-switch to revoke everything in one click.

🛡

Human Review Queue

Runtime violations show up as cards with full context: what the agent tried, why it was stopped, and what rule triggered. Approve, deny, or trust the pattern for 24 hours.

NEW
🌐

Egress Allowlisting

Lock your agent's network boundary to trusted domains only. Block outbound HTTP and webhook calls to unapproved URLs before they happen.

NEW
🧠

Thinking Token Sentinel

Monitors extended thinking usage (Claude 3.7+). If your model burns 50k+ thinking tokens without a tool call, Snapwire fires a “Potential Logic Loop” warning before costs spiral.

NEW
🔌

MCP-Compatible

Native JSON-RPC 2.0 support. Send standard MCP tools/call requests directly — Snapwire auto-detects the format. Works with Cursor, Replit, Claude, and any MCP-compliant agent.

NEW

Automatic Legal Protection

NISTIR 8596-aligned compliance reports, Safety PDFs, and Impact Assessments — generated automatically from your audit data. Tag high-stakes tools for Colorado SB24-205. Manage governance via API with Governance-as-Code.

SENTINEL
🛡

Zero-Code Runtime Security

Point your agent's traffic at Snapwire's Sentinel Proxy. No SDK. No code changes. Just swap your API base URL. 60 seconds to full governance.

# One line. Full governance.
OPENAI_BASE_URL=http://localhost:8080/v1 python my_agent.py

Supports OpenAI, Anthropic, MCP, A2A, and any JSON-RPC protocol. Three modes: Observe (silent audit), Audit (trace + log), Enforce (block + fail-closed).


Why Not Just a Sandbox?

Sandboxes protect your machine. Snapwire protects your business.

AI agent security has two layers. Most teams only have one.

📦

Layer 1 — Isolation

What sandboxes do

Keeps agents from touching your OS filesystem, host processes, or other containers. Essential for containment — but the agent still has full access to everything inside the sandbox.

Stops: filesystem escape, OS-level attacks, container breakout
You need this

Layer 2 — Governance

What Snapwire does

Intercepts what the agent tries to do — SQL queries, API calls, credential reads, spending — and enforces your rules before execution. Works inside any sandbox.

Stops: DROP TABLE, spending loops, credential leaks, policy violations

Snapwire runs inside any container, any cloud, any sandbox. The two layers are complementary — not competing.


Try It Now

Describe a rule. Get the code.

Type a safety rule in plain English. Snapwire generates the enforcement logic instantly.

Try:

FAQ

Frequently asked questions

Everything you need to know about Snapwire.

Snapwire sits between your AI agent and the tools it uses. Before the agent can call an API, modify a file, or spend money, Snapwire checks it against your rules and either approves it or holds it for your review.
Generate an API key, then send a POST request to /api/intercept before each tool call. Snapwire responds with "allowed", "blocked", or "pending". 60 seconds with any agent framework.
It goes into your review queue. You see exactly what the agent tried to do, why it was blocked, and which rule triggered. Approve or deny with one click. Unreviewed actions auto-deny after 30 minutes.
Sandboxes isolate the agent from your machine — they prevent it from touching your filesystem or OS processes. Snapwire operates at the tool call level — it intercepts what the agent actually does, like SQL queries, API calls, and spending, and enforces your business rules before execution. The two layers complement each other.
Yes. Create an organization to share rules, API keys, and review queues with your team. Roles (owner, admin, member) control who can change rules vs. who can only view.
Fork the repo on Replit or clone it and run with Docker — your own instance is live in 60 seconds. Snapwire is fully open source under the Apache 2.0 license, so you get the complete source code with no feature gates or usage limits.
Only ADMIN_EMAIL is required. DATABASE_URL defaults to a local SQLite database if not set, and SESSION_SECRET is auto-generated on first boot. LLM API keys (ANTHROPIC_API_KEY or OPENAI_API_KEY) are optional — add one to unlock AI-powered features like rule evaluation and drift detection.
Yes. All deterministic security features work without any LLM key — loop detection, spend tracking, schema validation, kill switch, Snap-Tokens, and honeypot tripwires are fully functional out of the box. AI-powered features like natural-language rule evaluation, drift detection, and tool grading require an API key.
Pull from the upstream repo using git merge or simply re-fork on Replit for a fresh start. Your rules and configuration can be exported and re-imported, so updates never wipe your setup.

Contact Us

Questions? Get in touch.

Need help with setup, have a feature request, or want to discuss enterprise deployment? We'd love to hear from you.

Your agents are already running. Start watching what they do.

60-second setup. Full spend protection and compliance evidence from day one. Catch your first runaway loop before it costs you.

Get Started Free

Shared Responsibility

You define the rules. We enforce them.

Snapwire provides the evidence-generation tools required for a "Reasonable Care" defense. The final configuration remains the sole responsibility of the Deployer.

Snapwire Does

  • Intercepts every tool call in real time
  • Generates audit logs with SHA-256 hashes
  • Flags high-risk actions for review
  • Produces compliance PDFs and Impact Assessments
  • Detects loops and spending anomalies
👤

Deployer Does

  • Sets risk thresholds and safety rules
  • Final approval of tool calls
  • Configures which tools are allowed
  • Reviews and acts on runtime violations
  • Engages qualified legal counsel for compliance

Snapwire is a technical monitoring utility. All blocks, alerts, and signals generated are heuristic and advisory in nature. The final Duty of Care for all agent actions and budgetary releases remains solely with the human operator.

Reasonable Care Disclosure: Snapwire's Sentinel Sidecar is a Deterministic Security Gateway designed to provide continuous monitoring and algorithmic guardrails for autonomous agents. By utilizing Snapwire, the Deployer implements a "Reasonable Care" framework to prevent foreseeable algorithmic harm. Snapwire serves as a technical control-plane for compliance with NIST AI Agent Standards.

Infrastructure Intermediary: Snapwire operates as a Passive Security Intermediary (Reverse Proxy). It does not generate, modify, or assume responsibility for the underlying intent or output of the AI Model. Snapwire's liability is limited to the integrity of the gateway's execution of user-defined rules. Forensic logs serve as the Immutable Forensic Record of intent and delegation.