Fleetwright MCP
Approved browser workflows, durable crawling and supervised recovery with inspectable receipts.

Deployed release · browser supervision and recovery verified
Independent build
An independent platform for owner-approved freight and maintenance workflows on owned destinations, plus catalogue crawling and a local agent. Arbitrary-site execution and commercial-scale availability remain unproven.
THE ENGINEERING
A deployed browser-operations platform with six scoped MCP tools, owner-approved freight and maintenance workflows on owned destinations, PostgreSQL fencing, a durable catalogue crawler and a local SQLite/Ollama agent. The October 14 correction adds browser supervision and readiness checks; the October 15 release publishes the operating, privacy, security and scale guides. Recorded regression passed 289 tests with six accounted skips; all 146 deployed artifact hashes and eight signed image identities matched. Actual owner booking, both approved MCP workflows, crawler export and all 15 CSP routes passed. A fresh archive restored into two new databases with matching workflow history, and scheduler/monitor recovery was exercised. Encrypted off-host restore and installed-model evidence retain their October 13 dates. Arbitrary-site execution, million-user capacity and measured monthly availability remain unproven.
BEFORE → AFTER
Real problems from this build, what I changed and the verified result.
- CASE 01Before
A worker can die after the destination accepts an action but before confirmation reaches the coordinator.
AfterThe owned-board contention evidence records 10,000 jobs and 200 coordination competitors with zero duplicate action attempts; competitors are not 200 browsers.
How I fixed itCombined current fencing, exact human approval and UNKNOWN states resolved from matching destination receipts.
- CASE 02Before
A healthy-looking worker can lack a usable browser driver and fail real bookings.
AfterThe corrected release passed fresh owner booking, both approved MCP workflows and a four-record crawler export. The October 15 regression passed 289 tests with six accounted skips.
How I fixed itAdded browser supervision and readiness checks, then exercised driver/browser loss, strict drain and reviewed recovery.
- CASE 03Before
A server-local backup and source hash alone do not establish recoverable deployment.
AfterAll 146 deployed artifact hashes and eight signed image identities matched. Both restored databases retained workflow history; encrypted off-host restore evidence remains dated October 13.
How I fixed itVerified signed image identities and exact deployed artifacts, then restored a fresh archive into two new databases and exercised supervised recovery.
Evidence reviewed: Source review October 15, 2026; dated verification and open gates stated above