● Independent boundary · Guided redirect · Routed action gates

Independent control
for AI agent actions.

Use any agent. Route risky actions through one policy boundary before they send, deploy, push, publish, or spend.

Show guided control flow Who it is for GitHub
Pre-runtime detection Redirects, not retry loops Bounded evidence
route ready
AgentVeil routed action-control flow A routed risky agent action reaches an action surface, passes through AgentVeil policy, redirects to a guided first step when needed, and records bounded evidence. Agent intent Surface action path Gate routed policy Redirect guided step Block no target Target routed Evidence bounded routed action routed gate post-action risk family guided first step target proof
Product thesis

The agent is not the control point. The action path is.

AgentVeil sits where agent actions leave the model and reach tools, APIs, repos, messages, deploy paths, packages, or production systems. Lurkr detects risky capability surfaces before runtime; AgentVeil redirects and gates routed actions when the agent tries to act.

Control flow

>Detect, redirect, gate, prove

Before runtime

Detect reachable risk

Lurkr checks what an agent could touch: tools, credentials, outbound sends, deploy paths, and risky capabilities.

→
During action

Redirect before target reach

Risky actions move to a guided first step: review diff, inspect package risk, review deploy plan, or review an external send.

→
Afterward

Gate and keep evidence

Routed actions can require approval or block, then preserve bounded evidence of the decision and target outcome.

Action surfaces

>Where the boundary matters

Built for teams shipping agents into real workflows, especially when agents can send messages, deploy, push code, publish packages, write data, or spend through tools.

SurfaceRisky actionSafe next pathBoundary note
Customer sendEmail, SMS, CRM, Composio-style sendsDraft, review, scoped approvalRedirect risky sends before customer contact.
Deploy / releaseWorkflow dispatch, release, package publish, paid API actionDeploy-plan review, staging, owner approvalRequire a routed control point before target reach.
Git / packagePush, publish, merge, package installPR, package review, credential custodyIf competing raw credentials exist, AgentVeil reports advisory.
High-stakes internalDB/API writes, internal console actionsMCP route, API gate, egress boundary previewMap what AgentVeil can and cannot enforce for one workflow.
Boundary truth

>Know what is protected and what can bypass

AgentVeil is not host-wide enforcement and not a Cursor, Codex, or Claude host lock. MCP controls only routed MCP traffic.

Stronger guarantees require a mandatory control point: credential custody, egress boundaries, routed remotes, or API gates. Until verified for a workflow, those are preview or design-partner boundary types.

Routed MCP actionprotected
Cursor / Claude connector pathprotected path
Standalone SDK/API routecustom route
Raw shell, visible credential, direct network pathbypass risk
Privacy and evidence

>Local where possible, bounded when shared

AgentVeil is built for sensitive agent workflows across the control loop. Local tools stay local by default. Hosted control APIs are designed around bounded metadata, hashes, and evidence rather than raw prompts, source code, secrets, or private logs.

What stays local

  • Lurkr scans locally.
  • MCP Proxy keeps raw MCP arguments, prompts, outputs, tokens, source code, secrets, and private logs local by default.
  • AgentVeil does not train on or sell customer data.

What becomes evidence

  • Bounded action metadata, risk class, resource hash, payload hash, policy context hash, timestamps, and target outcome.
  • Decision evidence that can be exported for clients or reviewers.
  • Proof that sits before the action, not just logs after the fact.
Start here

>Connectors and Quick Start

Start with a ready connector when your agent runs inside an IDE or agent runtime. Use standalone MCP Proxy when you want to wrap an MCP server directly. Use SDK/API surfaces for custom routed actions, approvals, evidence, and verification.

Cursor connector

Project-local control path for Cursor workflows.

Claude Code connector

Project-local control path for Claude Code workflows.

Standalone MCP Proxy

Manual route for MCP servers and stdio MCP clients.

Custom SDK/API

Custom routed actions, approvals, evidence, and verification.

# Cursor connector
pip install agentveil-mcp-proxy
cd <project>
agentveil-mcp-proxy setup cursor --yes
agentveil-mcp-proxy setup status --client cursor
# Claude Code connector
pip install agentveil-mcp-proxy
cd <project>
agentveil-mcp-proxy setup claude-code --yes
# restart Claude Code for this project
agentveil-mcp-proxy setup status
Standalone MCP Proxy commands 
# Manual route for MCP servers and stdio MCP clients
pip install agentveil-mcp-proxy
mkdir -p .agentveil
openssl rand -base64 32 > .agentveil/passphrase.txt
chmod 600 .agentveil/passphrase.txt
agentveil-mcp-proxy init --home .agentveil --quickstart-filesystem . --passphrase-file .agentveil/passphrase.txt
agentveil-mcp-proxy doctor --home .agentveil --passphrase-file .agentveil/passphrase.txt
agentveil-mcp-proxy client-config print --home .agentveil
POST/v1/runtime/evaluate
POST/v1/execute
GET/v1/execution/receipts
GET/v1/audit/verify

SDK: github.com/agentveil-protocol/agentveil-sdk · PyPI

>Adopt AgentVeil in stages

Start with one routed workflow in preview. Expand only when you need more boundaries, client proof, longer rollout support, private deployment, or higher operational limits.

PREVIEW

Build

Map one risky agent workflow, run Lurkr detection, and decide whether the path is protected, advisory, or unsafe.

  • ✓ Lurkr pre-runtime checks
  • ✓ Redirect playbook recommendation
  • ✓ MCP Proxy local setup
  • ✓ Bounded evidence + verification
  • ✓ 30-day retention during preview
Quickstart
EARLY ACCESS

Client Proof Pack

For SaaS and automation teams that need guided control for one sensitive agent workflow

  • ✓ Everything in Build
  • ✓ One routed agent workflow
  • ✓ Guided redirect guidance
  • ✓ Approval flow and bounded proof packets
  • ✓ Client verification link and proof export
Request →
PRIVATE ROLLOUT

Deploy

For teams expanding action control across production workflows and high-stakes environments

  • ✓ Everything in Client Proof Pack
  • ✓ Multiple workflows, custom thresholds
  • ✓ Private deployment options
  • ✓ Compliance & security review path
  • ✓ Retention, evidence export, support, and SLA
Discuss deployment →