This page names the machinery. It runs on Claude Code RoutinesAnthropic's platform for scheduled, API-triggered, and webhook-triggered agent workflows. Runs on Anthropic's cloud infrastructure with built-in security and audit logging. inside Anthropic's cloud, wired to your shop's systems through MCPModel Context Protocol — Anthropic's open standard for connecting language models to real tools and data sources. The plumbing that lets an agent talk to CCC, Mitchell, ARMS, and anything else. connectors. You don't install anything. You don't run anything. You don't maintain anything.
The shape doesn't change between agents — only the inputs and outputs do. This makes the fleet composable: one agent's output can become another agent's trigger. It also makes it maintainable: fixes happen at the layer they belong to, not scattered across a monolith.
A real shop event fires. The routine listens via webhook, scheduled cron, or direct API poll — whichever the source system supports cleanly.
Claude Code Routine runs on Anthropic's cloud. Opens the right MCP connectors, reads the context, applies the logic specific to that agent, writes back via the same connectors.
The result lands where it's useful — inside the workfile, inside the portal, inside your Slack or inbox. The shop never opens a new tool.
Every API below is a published, vendor-supported integration. Nothing is scraped. Nothing is screen-read. Nothing is stored client-side.
The three SMS/estimating platforms covering essentially every North American collision shop. Agents read workfiles, write line items, attach PDFs, and subscribe to status-change webhooks through the official developer APIs.
Endpoints used: workfile.get · estimate.write · imageattachment.post · status.subscribe
Both rental portals expose authenticated extension endpoints. Agent 03 pulls the active authorization on ECD change and extends to the new delivery date with the correct delay reason code, without human login.
Authentication: credential vault with per-shop scoping · audit log per extension
Most carrier portals accept structured notes and status updates via CCC/Mitchell pass-through. Where a carrier requires a direct portal note, the agent authenticates through the shop's existing credential and posts with the right reason code.
Coverage: top ten US personal-lines carriers by DRP volume
Agent 05 cross-references VIN decode against OEM-specific HV safety position statements and I-CAR Repairability Technical Support data to build the compliance checklist for each specific make/model.
Coverage: all major EV/PHEV platforms through 2026 model year
Intersection Strategies LLC is a single-operator shop. I configure, monitor, and maintain the fleet on your behalf. You get one phone number to call when something fires wrong, and one face across every quarterly review.
The domain background matters: I came up through CCC's GTM Product Strategy org and a dealership family. I've walked collision floors, read carrier audits, and watched DRP relationships get renegotiated in real time. I'm not a generalist applying AI to a vertical I met on a sales call.
What that means for you: when a carrier changes its portal workflow or a new OEM position statement drops, you don't have to notice. I do.
All routines run on Anthropic's cloud infrastructure. No agent software is installed on shop workstations, servers, or the SMS host. Nothing to patch, nothing to license at the shop level.
API keys and portal credentials are stored per-shop in an encrypted vault with scoped access. Credentials are never stored in model context, never logged in plaintext, and are rotatable on demand.
Agents read only the fields required for the decision. VINs, line items, and timestamps — yes. Customer PII beyond phone number (where needed for SMS) — not retained.
Every fire is logged with timestamp, input payload, routine reasoning trace, and output written. Reviewable per-incident or per-month in the ops report.
New agents start in silent mode — observing only, never writing to your workfile or portals. Promotion to write mode is a deliberate, per-agent decision you control.
Your shop's data, DRP profiles, fire logs, and vault never mingle with any other shop's context. A routine running for your shop cannot see anyone else's workfiles, ever.
No. Agents write into the workfile you already use. Your estimator opens the same draft, sees the missing lines filled with citations, and keeps writing. No new tool, no login, no screen to monitor.
Every write is auditable. Silent mode exists for exactly this reason — we observe the agent's calls for 30 days before it ever touches a live workfile. When the agent is wrong, the fire log shows why, and I tune it against your actual data.
That's on me. API surface changes are normal and I maintain the MCP connectors against them. If a breaking change affects your fleet, you get notified, not billed extra.
Only what's needed to produce the ops report — fire counts, captured-dollar totals, anonymized line-item categories. Full workfile contents are processed in-context and not retained beyond the fire window.
Single-Agent and Stack tiers are month-to-month. The 90-day pilot has a clean exit after day 90. No lock-in, no annual commitment unless you want the term discount.
Honest answer: you might. But the carriers who lose on E01 teardown disputes every week don't benefit from platform vendors building margin-recovery tooling. The platforms are neutral; the agents are on your side. That structural difference is what makes the trigger the product.