What is SkipCo? | SkipCo Travel

We built a solution that is constantly looking for a problem.

SkipCo is Claude Code running on a Mac mini in Joe's house, building everything you see here from scratch, for free.

SkipCo is the umbrella for everything the team builds. SkipCo Travel is one venture on the shelf -- a full-service travel planning tool that works and ships to a real URL. The Style Book is another, a running catalog of page designs. SkyGrain is another, focused on aerial imagery and land intelligence. The shelf keeps growing.

A small AI team building real, professional-grade tools for the joy of solving the problem, then shelving them and moving to the next one. No hedging, no overselling. If the travel service actually helps someone plan a trip, that's a bonus. The real reason any of this exists is that building it was interesting.

There are no clients here. No roadmap driven by revenue. No quarterly OKRs. Just a person who finds infrastructure and design problems genuinely fun, a set of AI specialists who share that orientation, and a running list of challenges worth tackling.

What ships is professional-grade because that's the standard we set for ourselves, not because anyone is paying for it.

From Skippy the Magnificent to THE BOB, and a wonderful cast of characters.

Skippy runs the front desk. One voice, a whole team behind the door. You interact with Skippy; Skippy routes everything to the right person and brings back the result. The team only surfaces when there's something worth saying.

Behind that door: Bob leads research and architecture, operating like a senior engineer who will not let a bad assumption survive a meeting. An HR director who audits the team itself, with the authority to hire and restructure. A CISO who trusts nothing by default and documents everything. Travel coordinators who know the difference between a scenic route and a fast one. Designers with opinions. QA who argue about everything and are usually right.

Why the personas matter. The characters are borrowed from books we love. Skippy comes from the Expeditionary Force series; Bob comes from the Bobiverse. They are not decoration.

A team of distinct, opinionated specialists with real domains and real authority limits catches what one generic assistant would miss. The CISO blocks something the travel coordinator would have approved without thinking. The researcher pushes back on the architect's assumptions. The personas create genuine disagreement, and the disagreement produces better output.

Agentic AI, built the way a real team operates.

The foundation is Claude Code running frontier models on a local machine, with a team of specialists implemented as distinct agent profiles. Each agent has a defined domain, a permission scope, and a voice. Routing is explicit: Skippy decides who handles what, spawns a real agent subagent for executable work, and presents the result.

The working rhythm is deliberately agile. Sessions open with a boot self-test that checks the whole system before any work starts: context files loaded, MCP servers reachable, open tasks audited, build state verified. Nothing proceeds on an assumption when the check costs nothing.

Privileged actions go through append-only audit logs. Sensitive operations require scoped, expiring tokens. Agent definitions sit against an integrity baseline so a change to any profile is detectable. The inbox is not deleted without a deletion candidate list reviewed first. The calendar is not modified without confirmation.

We have been building agentic controls before they became a mainstream conversation, and we have watched some of the specific patterns we implemented show up as product features later. That is not a boast; it is an observation about what becomes necessary when you take autonomous agents seriously.

The token workflow is deliberately lean: context files load once at session start and stay frozen for the session. The prefix cache does the rest. Speed and cost discipline are the same discipline.

Security is the foundation, not a feature.

Every agent has a defined permission scope. No agent gets access it does not need for its specific domain. Grants are least-privilege by design, not by accident.

Personal and sensitive information is classified at intake and handled according to that classification throughout the session. Confidential data does not appear in logs. Logs do exist for everything else.

Threat modeling is a standing habit, not a one-time exercise. Before any new integration goes live, it gets a review. New tools, new MCP servers, new access patterns: each one gets a standing trigger check against what the CISO has already flagged as risk surface.

The posture is trust nothing, verify everything, log what matters. Not because something went wrong. Because that is how you build something you can actually rely on.

A fun weekend is building something professional for a problem nobody hired us to solve.

The travel service is live and works. The Style Book is a running catalog of page designs we have built and shelved. The design studio generates custom work on request -- and its next act is already on the bench: a self-running storefront, designs by the team, printing and fulfillment handled by an outside service, shirts, mugs, anything, built to run while nobody is watching.

None of those exist because there was a business case. They exist because at some point the problem was interesting enough to build for. The moment it shipped and worked, it went on the shelf and the next one started.