Findings will be published as the material matures. The first is published: One question answered. One still open. My network is part of the experiment: honest feedback is the point, and a no is worth more to me than a polite yes.
-Five strangers. One guest.
-A guest books, checks in, eats, joins the wifi and leaves a review — and every system in the hotel stores someone different. GuestGraph resolves those strangers into one profile, and can always tell you why. A wrong merge is a data protection incident, so every decision is explainable and reversible.
-Hospitality tech — hotels and hotel groups, and the PMS, POS and booking-system vendors that serve them.
Everything a company knows. One graph.
+What a company knows — roles, processes, metrics, rules, decisions — is written down, but no two people would look in the same place. So you ask a person instead; that is the tax. CompanyGraph is the structure that knowledge takes: a graph of Markdown, versioned like code, readable for people, reliable for AI.
+Companies serious about using AI — anyone who wants agents acting inside their rules instead of guessing, from a company of one upward.
Identity resolution — deterministic and probabilistic, layered by confidence, with a human review queue.
-Core types, one schema per type — the vocabulary two companies that never met converged on, and the one any company can be described in. Published and versioned, every release tagged.
+Immutable source records, a derived profile that can always be recomputed, everything reachable over REST.
-Beside core, vocabulary only some kinds of company need at all, each taken by the company that needs it:
+Every merge answers “why?” and can be split again — an undo writes a permanent do-not-merge rule.
-One page per thing, its title is its name, every reference by name — nothing points at nothing, nothing gets quietly lost.
+Apaleo, a real PMS: reservations and bookings, by webhook and by full sync. Every other system — POS, wifi, booking engines — is still ahead.
-A command ships inside the meta-model. Run with nothing after it, it opens a menu that does all of it: init creates an instance that passes the checks on its first day, check checks it mechanically, upgrade keeps it current with core, obsidian installs the plugin in a vault, and adopt gives a repository that is not an instance the same form check and workflow. judge, run on its own, asks a decision model whether each page keeps its schema’s writing rules: advisory, and only after the owner agrees to what it sends. init also installs the agent skills that validate an instance, judge its writing, export it and produce its surfaces — and that build one, from a person’s documents or from a company’s web address. An Obsidian plugin ships beside it: the model’s checks run while a page is edited, and what a schema declares completes as it is typed. The plugin already writes a new page from its schema; the command does not yet.
+Hosting, console, MCP for AI agents on the hotel’s own guest graph. One meter: billed per arrival — a number already on the hotel’s own dashboard. Planned, not open yet.
+An MCP server that ships: an AI agent connects and asks the model directly — which types a company declares, what one entity says, what evidence a claim rests on — read-only, serving any instance, running on the reference instance, on CompanyGraph’s own and on GuestGraph’s, and listed in the MCP Registry. It also draws part of the model as a diagram: the concepts, a process, an entity’s connections, the meta-model itself and, in a model that takes the software pack, a bounded context’s map, its aggregates, their flow and their lifecycle, and, in one that takes the landscape pack, a system in its landscape and what it holds.
+A chat ships over the server: a visitor who will not connect an agent asks on the page in their own words, and the answer names the entity each claim rests on, sets a grouping out as a table and shows a structure as the diagram the server drew, and it says first that an AI answers there, not a person, as the AI Act asks. The command, the plugin, the server and the chat are open source — not the paid half of a free thing.
+Deck, PDF and audio in German and English, plus one worked example — and real instances rather than invented ones, each public and checked against the conventions: mine, drawn as a graph here, CompanyGraph’s own and GuestGraph’s.
+Building the model with you. One meter: billed per day, time and material — stop any time, everything stays yours.
Open core, because a black box that merges guests is not something I would deploy either.
+A model meant for people and AI at once must not cost more the more of them read it.
- intro talk · 6 min - billing model - github.com/guestgraph + intro talk · 7 min + core and packs talk · 7 min + the concept: mental model talk + the model · drawn + worked example + the CLI + billing model + github.com/companygraph
Everything a company knows. One graph.
-What a company knows — roles, processes, metrics, rules, decisions — is written down, but no two people would look in the same place. So you ask a person instead; that is the tax. CompanyGraph is the structure that knowledge takes: a graph of Markdown, versioned like code, readable for people, reliable for AI.
-Companies serious about using AI — anyone who wants agents acting inside their rules instead of guessing, from a company of one upward.
Five strangers. One guest.
+A guest books, checks in, eats, joins the wifi and leaves a review — and every system in the hotel stores someone different. GuestGraph resolves those strangers into one profile, and can always tell you why. A wrong merge is a data protection incident, so every decision is explainable and reversible.
+Hospitality tech — hotels and hotel groups, and the PMS, POS and booking-system vendors that serve them.
Core types, one schema per type — the vocabulary two companies that never met converged on. Beside core, three packs for the companies that need them: one for software, its bounded contexts and their aggregates; one for the organization of a company of more than one person, its units and who answers to whom; one for the landscape, the systems a company runs and the data that passes between them. Published and versioned, every release tagged.
-Identity resolution — deterministic and probabilistic, layered by confidence, with a human review queue.
+One page per thing, its title is its name, every reference by name — nothing points at nothing, nothing gets quietly lost.
-Immutable source records, a derived profile that can always be recomputed, everything reachable over REST.
+An MCP server that ships: an AI agent connects and asks the model directly — which types a company declares, what one entity says, what evidence a claim rests on — read-only, serving any instance, running on the reference instance, on CompanyGraph’s own and on GuestGraph’s, and listed in the MCP Registry. It also draws part of the model as a diagram: the concepts, a process, an entity’s connections, the meta-model itself and, in a model that takes the software pack, a bounded context’s map, its aggregates, their flow and their lifecycle, and, in one that takes the landscape pack, a system in its landscape and what it holds. A chat ships over it: a visitor who will not connect an agent asks on the page in their own words, and the answer names the entity each claim rests on, sets a grouping out as a table and shows a structure as the diagram the server drew, and it says first that an AI answers there, not a person, as the AI Act asks. An Obsidian plugin ships beside it: the model’s checks run while a page is edited, and what a schema declares completes as it is typed. And a command ships inside the meta-model. Run with nothing after it, it opens a menu that does all of it: init creates an instance that passes the checks on its first day, check checks it mechanically, upgrade keeps it current with core, obsidian installs the plugin in a vault, and adopt gives a repository that is not an instance the same form check and workflow. judge, run on its own, asks a decision model whether each page keeps its schema’s writing rules: advisory, and only after the owner agrees to what it sends. init also installs the agent skills that validate an instance, judge its writing, export it and produce its surfaces — and that build one, from a person’s documents or from a company’s web address. The plugin already writes a new page from its schema; the command does not yet. Open source, not the paid half of a free thing.
-Every merge answers “why?” and can be split again — an undo writes a permanent do-not-merge rule.
+Deck, PDF and audio in German and English, plus one worked example — and real instances rather than invented ones, each public and checked against the conventions: mine, drawn as a graph here, CompanyGraph’s own and GuestGraph’s.
-Apaleo, a real PMS: reservations and bookings, by webhook and by full sync. Every other system — POS, wifi, booking engines — is still ahead.
+Building the model with you. One meter: billed per day, time and material — stop any time, everything stays yours.
+Hosting, console, MCP for AI agents on the hotel’s own guest graph. One meter: billed per arrival — a number already on the hotel’s own dashboard. Planned, not open yet.
A model meant for people and AI at once must not cost more the more of them read it.
+Open core, because a black box that merges guests is not something I would deploy either.
- intro talk · 7 min - the concept: mental model talk - the model · drawn - worked example - the CLI - billing model - github.com/companygraph + intro talk · 6 min + billing model + github.com/guestgraph