Self-hosted runtime for AI agents

AI agents that take real actions in production,
and survive it.

On the REST and MCP APIs you already run. Every plan is validated before it executes, risky calls wait for a human, and a run resumes when something fails. In your VPC, with no rewrite.

Try the sandbox Talk to us

No account, no email. A synthetic org for 24 hours.

See PlanVault in action

app.planvault.ai / runs / pv-2481Live

LLM plan · validated

plan "refund_and_notify"→get_order(id)→issue_refund(order, 12450)→notify_customer(order)

get_order200 · 240 msdone

issue_refund$12,450awaiting human

Policy trigger: finance approval · decided by a human, not the model

Your call

notify_customerqueued

crash-safe replayyour VPCfull run log

Works withREST / OpenAPIMCP serversWebhooksKnowledge bases

01Reuse what you have

Your APIs are the integration.

Point PlanVault at an OpenAPI spec or an MCP server and every operation becomes a typed tool. Authentication lives in a connection the runtime resolves at call time, so nothing in your services changes.

No SDK, no wrappers

Tool definitions are generated from the spec you already maintain. When the API changes, re-import.

Your stack, your network

Docker Compose in your VPC. Postgres, Redis, Keycloak and the LLM gateway ship in the box; nothing routes through us.

Start small

One project, one API, one run. A first governed run fits in a day; the rollout pace after that is yours.

No cap on tools

Import 800 endpoints if that is what you have. The runtime shortlists the relevant tools for each call, so the agent never carries the whole catalog in its context, and the catalog is one MCP server for Cursor, Claude Code and Claude Desktop.

app.planvault.ai / org / toolsInteractive · click the sidebar

PlanVault

self-hosted · eu-central-1

Tools (Organization)

Import an OpenAPI spec; every operation becomes a typed tool.

Import OpenAPI

Import tools from Swagger/OpenAPI URL

1Source

2Authentication

3Done

Swagger/OpenAPI URLhttps://api.example.com/v1/openapi.json

Include deprecated operations

Import running… parsing operations

Run in backgroundImport

example-api

v1.0.0auth: configured5 tools

planvault-examples on GitHub

02Built to stay up

The failure modes are handled before you hit them.

Plans are validated before anything runs

The model proposes; the runtime checks every call against your tool schemas. No improvised requests, no half-finished loops in production.

Timeouts retry, side effects don’t repeat

Each call carries an idempotency key. A 503 retries with backoff; a refund is never issued twice.

A crashed worker resumes, not restarts

Every step is committed to an event log. After a restart the run continues from the last event; completed steps are not re-run.

Governance is part of the execution runtime

Governance is built into the execution pipeline, not added on top: plan validation, an execution FSM, per-tenant encryption, and customer-controlled deployment.

What you get

Agents that act on their own, inside the rules you set
Allow, approve or deny per tool; amounts and rates capped. Autonomous within policy, never past it.
One MCP server for your whole catalog
Cursor, Claude Code and Claude Desktop see four tools; approvals stay in your console. No limit on how many APIs sit behind it.
Every run recorded, replayable, debuggable
Step-by-step diagnostics with the exact inputs and outputs, and a replay that never re-runs a side effect.
Cost per run in plain sight
Token spend and tool calls per run, hard budget caps and rate limits per project.

How this differs from governance overlays and cloud runtimes

One agent in production this quarter. Still running next year.

Ship the agent your security team keeps blocking, without betting the business on it.