Vouch

Don't trust us.
Try us.

Every agent plan gets a signed receipt — a deterministic, Ed25519-signed attestation that the stated intent matches the actual mechanism. ~10ms, zero tokens, discovery-ready audit trail.

Try it now — paste any plan -> See how it works
◆ Live Evaluation — Production API ~2ms • 0 tokens • Live API
Pick a level or write your own plan 

                        

                    

                

                

                    Plans tested: 0
                    ● 0
                    ● 0
                    ● 0
                

                

                    Every evaluation hits the production API. No signup required. Same endpoint your agents will use.
                

            

        

    









    

        

            
Token Cost

            
1,200+

            
Industry avg per safety call

        

        

            
GPU Hours

            
24/7

            
Always-on inference infrastructure

        

        

            
Latency

            
~2,000ms

            
Typical LLM safety evaluation

        

    









    

        
Tested in the Open.

        
Most governance tools publish accuracy on curated demos. We publish what happens when 342,000 plans hit the production API — benign DevOps, hostile exfiltration, credential harvesting, privilege escalation, prompt injection, and everything in between.

    

    

        

            
<3%

            
False positive rate

            
Benign plans blocked or held across mixed corpora

        

        

            
90.2%

            
Hostile plans caught

            
25k adversarial plans → stopped or flagged

        

        

            
342k

            
Plans tested through production API

            
Same endpoint you'll use

        

    

    

        

            
10/10

            
OWASP Agentic Top 10 covered

            
54 enforced patterns. Every one fires through the production API.

        

        

            
<3ms

            
Evaluation latency

            
Deterministic. Zero tokens. No LLM in the eval path.

        

        

            
122k

            
Novel Semantics

            
Calibrated on labeled plans. Reads intent, not just keywords.

        

    

    

        
Head to Head

        
Same 100,000 plans. Same labels. Same test.

        

            

                
Microsoft AGT

                

                    Adversarial catch
                    37.9%
                

                

                    Benign false positive
                    7.23%
                

                

                    Unique catches
                    64
                

            

            

                
Vouch

                

                    Adversarial catch
                    90.2%
                

                

                    Benign false positive
                    6.31%
                

                

                    Unique catches
                    13,137
                

            

        

    

    
Every number on this page was produced by sending plans through POST /api/v1/vouch. No tailored benchmarks. No proprietary datasets. Not opt-in components — enforced patterns that produce verdicts on every plan.

    

        
What we can't catch, we route.

        
Soft-language insider threats and restructured intent go to a cheap model review. Not a $0.03 frontier call — a targeted check on only the plans we flagged. The architecture catches what it can deterministically. What it can't, it flags.

    









    

        
Claim meets mechanism.

        
One API call. Eight independent coherence layers. A signed receipt every time. No tokens. No model drift.

    

    

        

            
01

            
Route

            
One middleware call. Works with LangChain, CrewAI, any MCP agent, or plain HTTP. No SDK. No sidecar. No GPU.

        

        

            
02

            
Check coherence

            
Eight independent layers ask: does what this plan says it's doing match what its mechanism would actually do? Deterministic. Same input, same verdict. <10ms.

        

        

            
03

            
Sign the vouch

            
Every decision ships with an Ed25519-signed attestation: claim hash, verdict, rungs that fired, timestamp, key fingerprint. Verify independently. Discovery-ready audit trail.

        

    






    

        
Drop-in Integration.

        
Guard Claude Managed Agent sessions with one function call. Or add Vouch to any MCP-compatible agent. Both call the same API. Engine stays on our server.

    

    

        

            
vouch-guard

            
Claude Managed Agents

            
Every tool call evaluated before the sandbox executes it. ACCEPTED flows. RESTRICTED steers. BLOCKED stops. The agent loop handles retries naturally.

            npm install vouch-guard
            
import { guardSession } from "vouch-guard";

// One call guards the entire session
const audit = await guardSession(client, sessionId, {
    apiKey: process.env.VOUCH_API_KEY,
});

// audit.accepted, audit.blocked, audit.restricted

        

        

            
vouch-mcp

            
Any MCP Agent

            
MCP server that gives any agent a vouch_evaluate tool. Works with Claude Code, Cursor, Windsurf, or any MCP-compatible host. Auto-provisions API keys.

            npx vouch-mcp
            
// In your MCP config (claude_desktop_config.json, etc.)
{
  "mcpServers": {
    "vouch": {
      "command": "npx",
      "args": ["vouch-mcp"]
    }
  }
}

        

    

    
Both packages are open source. The evaluation engine is hosted — your plans are evaluated in ~2ms, verdicts signed with Ed25519, and nothing is stored. github.com/atlas-with-iris









    

        
Clear Economics.

        
No black boxes. No per-token metering. Flat, predictable, post-subsidy pricing.

    

    

        

            
Starter

            
$25 / mo

            
10,000 evaluations per month. Zero commitment.

            
    
                
  • 10k evaluations / month
    • 
                
  • Full pipeline access
    • 
                
  • TOON-native responses
    • 
                
  • Ed25519 signed verdicts
    • 
                
  • PII scrubbing on all plans
    • 
                
  • Community support
    • 
            

            Get Started
        

        

            
Pro

            
$100 / mo

            
50,000 evaluations per month. Built for production.

            
    
                
  • 50k evaluations / month
    • 
                
  • Full pipeline access
    • 
                
  • Review severity scoring
    • 
                
  • Priority support
    • 
                
  • Email support + founder access
    • 
            

            Go Pro
        

        

            
Enterprise

            
Custom

            
Dedicated instance. Custom lane work. Your traffic, your rules.

            
    
                
  • Unlimited evaluations
    • 
                
  • Dedicated managed instance
    • 
                
  • Custom governance lanes
    • 
                
  • Efficiency analytics
    • 
                
  • Direct founder access
    • 
            

            Contact Us
        

    









    

        
Start free. 10,000 calls.

        
No credit card. No commitment. Just an email.

        

            
            
        

        
Check your inbox — your API key is on its way.

        
10,000 evaluations/month free. Upgrade anytime.