OPERATOR by rytstack
Let AI agents safely operate legacy web apps - no APIs required
Your AI agents need to act in web apps that have no usable API. OPERATOR gives them reviewed, deterministic, drift-monitored tools, delivered as portable MCP servers that run behind the gateway you already use.
AI agent
Any agent framework
Your MCP gateway
Identity, policy and audit stay with you
OPERATOR
Compiled, reviewed MCP tools
Legacy web app
No API required
The problem
AI agents need to act where there is no API
Enterprises need AI agents working in legacy web apps. The usual options each fall short.
Bespoke API or connector
Rarely funded for a single web apps, so the work never starts.
RPA platform
Heavy for a handful of apps, with its own runtime, licences and maintenance burden.
Generic browser agent
Takes a different route each run. Hard to audit, hard to certify and exposed to prompt injection from page content.
How it works
From web apps flow to production tool
The LLM works at authoring time only. What reaches production is compiled, versioned and reviewed.
Step 1
Record
Walk through the web app flow once, or let OPERATOR crawl it. An LLM drafts a task-level tool.
Step 2
Review
A person checks the name, typed schema and risk class, then publishes a versioned tool.
Step 3
Run
The tool replays the web app’s own network calls, with browser steps as the fallback. No LLM at run time.
Step 4
Monitor
Scheduled read-only checks catch web app changes. Affected tools are taken offline until repaired.
The platform
Everything between your agent and the web apps
Authoring tool
Record or crawl a flow. An LLM drafts the tool and a human reviews and publishes it.
Compiled executors
Deterministic replay of the web app’s network calls, with browsing steps as the fallback.
Tool registry
Versioned tools, each tagged read, reversible write or irreversible write. Export as an MCP server.
Credential and session broker
Holds web app credentials and sessions so agents never see them.
Drift monitor
Scheduled synthetic calls and contract tests. Changed tools are taken offline and a repair is proposed.
Gateway integration
Structured audit events to your SIEM. Identity and policy decisions come from your gateway.
Designed for regulated teams
Deterministic, reviewable and auditable
Web app content is untrusted and every action has a risk class. That keeps the agent’s blast radius small and the evidence trail clear.
- Task-level tools only
- Small tool sets per agent
- Typed fields, not raw page text
- Irreversible actions flagged for approval
Reviewed before release
Every tool is versioned and approved by a person. Nothing changes on a regulated process silently.
Resistant to prompt injection
Typed extraction drops free text where possible, and output is validated before it reaches the model.
Safe writes
Idempotency keys and a check-before-submit step prevent duplicate claims, demands and payments.
Credentials stay out of prompts
The broker holds secrets and sessions, so they never appear in prompts, config files or logs.
Evidence on demand
Structured audit events show who did what and why, ready for your logging or SIEM.
Quick to contain
Take offline a single tool or roll back a version without taking the whole estate offline.
Is OPERATOR right for you?
Built for teams with legacy web apps
Who it is for
Enterprises with web apps their agents must use. Integrators who rebuild these connectors for each client.
Works with your stack
Portable MCP servers run behind your gateway and any agent framework, on AWS, Azure or open-source tooling. OPERATOR supplies the tools and your gateway enforces policy.
Not in scope today
Desktop apps, Citrix, mainframe terminals, and apps protected by CAPTCHA or device binding.
Questions
Frequently asked questions
Do you replace our MCP gateway?
No. Gateways already handle identity, policy and audit well. OPERATOR supplies the tools behind them, emits audit events to your logging, and accepts policy decisions from your gateway.
Does an LLM run every time a tool is called?
No. The LLM drafts a tool at authoring time. A person reviews it, and the published tool then runs deterministically.
What happens when a web apps changes?
The drift monitor runs scheduled read-only calls and contract tests. If a schema or page changes, the affected tool is taken offline, an alert is raised and a repair is proposed for human review.
Is it permitted to automate our apps?
That depends on each web app’s terms and your contracts. We limit tools to apps you are licensed to automate and recommend a legal review for each one.
Where does it run?
OPERATOR is delivered as portable MCP servers, with a customer-hosted deployment option for regulated environments.
Have web apps your agents cannot reach?
Join our Design Partner Program: Bring your toughest internal web app, and our team will compile your first production MCP tool pack in a live working session.