Overview
Preloop is an Apache-2.0, self-hostable control plane for AI agents. It combines model traffic management, MCP tool governance, human approval workflows, runtime visibility, and audit records without requiring SDK or agent code changes.
It supports coding agents and MCP-compatible runtimes, including Claude Code, Codex CLI, Cursor, Gemini CLI, Hermes, OpenClaw, OpenCode, Windsurf, Copilot CLI, and standard MCP clients.
Key Features
- Routes OpenAI- and Anthropic-compatible model traffic through a gateway that attributes usage to agents, flows, or API keys.
- Applies soft and hard budgets per account and flow while marking usage it cannot price as unpriced.
- Governs MCP tool calls with ordered YAML policies and CEL expressions that can inspect arguments and context.
- Allows, denies, or pauses calls for approval through mobile, watch, Slack, Mattermost, email, or webhooks.
- Records tools, inputs, matched rules, decisions, approvers, model spend, and outcomes in session audit trails.
Agent Onboarding
The Preloop CLI can discover compatible local agent configurations and report their capabilities. During onboarding, it imports supported MCP servers and model metadata, creates a credential, backs up the existing configuration, and rewrites the agent to use Preloop-managed firewall and gateway endpoints.
This approach lets teams place existing agents under common controls without modifying their source code. Allowed actions continue immediately, denied calls fail with an actionable message, and approval-gated actions can wait synchronously or use an asynchronous polling mode.
Governance and Deployment
Policies can cover deployments, shell commands, database operations, secret access, cloud provisioning, billing changes, ticket automation, and internal APIs exposed as tools. Rules remain alongside infrastructure configuration and can be version controlled.
Self-hosted deployments keep tool calls, model traffic, and audit records within the operator's network. The open-source core supports one operator per account, while hosted and commercial self-hosted editions add multi-user, team, and role-based access features.
Operational Evidence
Preloop can collect machine evidence for oversight and security programs by retaining action histories and approval decisions. Its Apache presets can verify a build SBOM, map components against vulnerability sources, and produce a versioned result.
These records may support work related to the EU AI Act, Cyber Resilience Act, DORA, and NIS2, but they are operational evidence rather than certification or legal advice.
