Preloop

Preloop

The open-source AI agent control plane and software factory.

Preloop is one self-hostable service that sits between your AI agents and everything they reach, and runs agents for you when work arrives. Every tool call and every model call is governed, attributed and visible, whether a person started the agent or a ticket did. You see gateway spend by model, agent, session, API key and flow, and cap it with budgets. Imported Copilot usage is displayed separately as vendor-reported data.

It has two halves, and the same policies, budgets and session timeline apply to both:

• Control plane. An MCP firewall that allows, denies or holds any tool call for human approval; an AI model gateway with budgets and attribution; policy-as-code in YAML and CEL; one runtime timeline per session.
• Software factory. Flows run agents, in hosted containers or on your own runners, when GitHub, GitLab, Bitbucket Cloud, Jira, a webhook or a CI job fires. Presets review pull requests, implement issues, triage tickets, scan for vulnerabilities and collect evidence packs. A human can steer a running agent from the console, phone or watch, and the work lands as pull requests you can review.

Both halves are multi-user: invite colleagues and organize teams in Preloop (OSS). Each user can connect devices, runners and agents. Preloop (OSS) has no roles or access controls between users of an account; those are in Preloop Cloud and Preloop Enterprise.

Who it is for:

• Platform and security teams that need agents governed before they reach production systems.
• Engineering teams running coding agents that want to see the cost and keep an audit trail.
• Teams that want tickets and events turned into bounded agent work they can review.

MCP clients connect to /mcp; existing /mcp/v1 configurations remain compatible. See the tested MCP protocol matrix for modern discovery, legacy initialization, and SDK migration details.

Works with OpenClaw, Claude Code, Codex CLI, Cursor, Gemini CLI, Hermes, OpenCode, Devin, Copilot CLI, and any MCP-compatible agent.

Your first five minutes

# 1. Install the CLI (macOS / Linux / WSL)

curl -fsSL https://preloop.ai/install/cli | sh

Windows, native (PowerShell): irm https://preloop.ai/install/cli.ps1 | iex

Details: docs/windows-cli.md

2. Connect it to a control plane

preloop signup # Preloop Cloud (fastest), or preloop login --url http://localhost:3000 # your self-hosted instance

3. Bring local agents under governance

preloop agents discover

preloop agents discover finds local agent configs and reports what each adapter supports. preloop agents onboard configures them: calls routed through the managed MCP entry reach the MCP Firewall, supported model routes use the Gateway. To talk to a running agent, install the runtime plugin with preloop agents install-plugin (or preloop claude for Claude Code). Full walkthrough: onboard local agents.

Watch it work

Onboarding, the MCP firewall, human approvals, and cutting session cost. Recorded against a real stack, no slideware.

Watch the full playlist →

What you get

Jobs teams otherwise buy from several vendors, in one Apache 2.0 stack:

AI Agent → Preloop → [Policy]  → Allow / Deny / Require Approval → Execute

→ [Gateway] → Budget + attribution → Model

The software factory

One loop, end to end: an issue event arrives from GitHub, GitLab, Bitbucket Cloud or Jira. A flow starts an agent in a hosted container or on your own runner. Every tool call it makes passes the same firewall and approvals as your own agents, and every model call is charged to the flow's budget. The agent opens a pull request; the Pull Request Reviewer reviews it; review or CI feedback resumes the implementation on the same branch, and the result is a pull request you can review. You can steer the run with an operator note at any point, and the whole run is one timeline.

21 presets ship in this repository, including:

• Issue Triage Assistant and issue refinement
• Pull Request Reviewer
• Delivery Report: a weekly report with proposals for human decisions and project reruns. See what landed, what the available evidence supports against stated criteria, and which gaps deserve investigation
• Automated issue implementation
• Security audit presets: SBOM verify, exploit check, release audit. They collect machine evidence for CRA- and EU AI Act-style reviews. That is not a conformity assessment, certification, or legal advice.
• Observe / Eval
• Component due diligence, architecture and strategy review, repo code health review
• Standards compliance walk, release backport, weekly model price review

Or write your own. Trigger from CI with the run-flow GitHub Action; run agents on private runners; let a flow start child flows with flow delegation. Tracker setup: Bitbucket and Jira quickstart; Bitbucket Data Center is available as an opt-in provider.

Policy-as-code

version: "1.0"

metadata: name: "Production Safeguards"

approval_workflows:

• name: "deploy-approval" timeout_seconds: 600 required_approvals: 1 async_approval: true

tools:

• name: "bash" source: mcp approval_workflow: "deploy-approval" justification: required conditions:
• expression: "args.command.contains('deploy') && args.command.contains('production')" action: require_approval

Check a sample call before saving with Test this rule or Simulate in the console. Ship it with preloop policy apply (validate / diff / export also exist).

More capabilities

• Model providers: OpenAI, Anthropic, Google, Amazon Bedrock, Azure OpenAI, Alibaba Cloud Model Studio, DeepSeek, Mistral, Moonshot, Z.ai, OpenRouter, any OpenAI-compatible endpoint, and runner-proxied local models. Reviewed model pricing; unpriced models stay visibly unpriced.
• Restricted CI identity: a machine identity bound to one project and one flow for CI jobs, with an optional read-only native GitHub App clone mode and separate result publication. Guide.
• Account kill switch: blocks gateway and tool traffic and stops managed flow runs, with audited recovery. Guide.
• Durable audit delivery: opt-in transactional intents for approvals, policy decisions and tool outcomes, with restart replay and backlog alerts. Operations guide.
• Session search: keyword search by default, semantic as an opt-in, every search audited. Saved searches.
• DORA exports: the AI-agent slice of the ICT inventory and incident record. Guide.
• OTLP export: off by default; emits OpenTelemetry GenAI spans. Guide.
• Chat: Slack, Mattermost, and Discord for questions, approval votes and operator notes. Chat connections.
• Event-driven employees: managed agents doing bounded tasks from tracker and error events. Guide.
• Remote sessions and deployment: [remote sessions](docs/guide/runners/remo