Agents that work
inside any system.
When a system has no API, Reach steps in. It controls real browsers — filling forms, clicking buttons, extracting data — with a five-tier risk model that keeps humans in control of every action that matters.
When to use Reach
MCP when you can. Reach when you must.
Structured integrations via MCP are always preferable — faster, more reliable, easier to test. Reach is the fallback for the real world: systems without APIs, portals that won't share credentials, legacy software your vendor stopped maintaining in 2015.
Use MCP when
- A native API or connector exists
- The system supports webhooks or event streams
- You need real-time sync at high volume
- The integration is business-critical and permanent
Use Reach when
- No API exists or is too expensive to build against
- The system is a legacy web portal
- The workflow is low-frequency but still needs automation
- The task is inherently visual (screenshots, PDFs, UI-driven reports)
Safety model
Five-tier action risk model
Not all browser actions carry the same risk. Reach classifies every action before executing it. The agent cannot skip a tier or misclassify an action — the safety layer is enforced at the tool wrapper level, below the LLM.
Read page content, screenshots, DOM inspection. Zero side effects. Always safe, never requires approval.
Click links, open URLs within allowed domains. Scope-locked — requests outside allowed URLs are blocked at the network level.
Type into fields, select options. On systems that auto-save (Notion, Salesforce, ERPs with no Save button) — requires pre-approval before the first keystroke.
Click Submit, Save, Confirm, or Send. Also triggers on Enter/Return in a form field. Always requires human approval — the agent pauses and waits.
Delete, archive, or irreversible actions. Maximum gate — requires explicit approval with a stated reason. Agent cannot proceed without it.
Agent requesting approval — submit action
About to click Login on vendor-portal.example.com. This will authenticate and begin the session.
How it works
Built for the real world, not the demo.
Human approval gate
The agent pauses before any risky action and presents a screenshot of exactly what it is about to do. You approve or reject in real time — no action executes without your go-ahead.
Scope enforcement
Allowed URLs and domains are declared when a session starts. Playwright request interception blocks every navigation outside that scope — the agent physically cannot leave the perimeter you set.
Auto-save awareness
Modern systems save on field change or blur — no Save button required. Reach detects auto-save and blur-save semantics and applies fill-level approval rules before the first keystroke, not after.
Screenshot audit trail
Every action in every session is logged with a before screenshot, the proposed coordinates, the action taken, and a success flag. Six months later you can reconstruct exactly what the agent did and why.
Session isolation
Each task runs in a dedicated Chromium instance with no shared state. Cookies, sessions, and local storage do not leak between tasks or tenants.
Dry-run mode
Run in dry-run before giving the agent real credentials. The agent navigates, observes, and plans — but every fill and submit requires approval, even low-risk ones. Review the plan before it executes.
Real use cases
What eClips Reach handles today.
Legacy procurement portals
A vendor portal from 2009 with no API and no SFTP export. Reach logs in, downloads the pending POs, extracts line items, and posts them to Odoo — without a human touching the keyboard.
Government e-services
ZATCA, Etimad, GOSI — portals that exist only as web forms. Reach submits filings, checks statuses, and downloads certificates on a schedule. Approval gates fire before anything is submitted.
Supplier self-service portals
A 3PL partner portal that won't share API keys. Reach reads shipment statuses, marks deliveries confirmed, and triggers the ERP update — with a screenshot of every confirmation screen saved to the audit log.
HR and payroll systems
A regional payroll platform with no export API. Reach navigates to the report, changes the date range, downloads the CSV, and hands it to the data pipeline. Every step logged. No human clicks required.
Agents that cannot go rogue.
The safety layer in Reach operates below the LLM. The agent cannot misclassify an action to avoid approval — classification and approval gates run in the orchestrator wrapper, not inside the model context. The browser's network requests are intercepted at the Playwright level, so scope violations are physically impossible, not just discouraged.
Every session is tagged with the scope configuration, the user who launched it, and the approval decisions made during the run. If something goes wrong, you have a complete chain of custody.
Which system are you trying to integrate?
Tell us about the portal, legacy system, or web form your team handles manually. We'll tell you whether Reach can automate it and what the approval model would look like.
Your Business,
On Autopilot.
From project management to quality systems to security — we build intelligent platforms that simulate, automate, and transform how your operations run.
30 minutes · No pitch deck · Live product walkthrough