infinimind creations

The lab

Machines

Everything on this site runs on machines that stand in one room. Nothing here is rented by the hour. Four boxes, sixteen guests, a git home, a vault, and one set of documents that describes all of it.

4physical machines
16virtual machines running, of 17 defined
50repositories in the lab's own git

The machines

Four boxes, sixteen guests

3 virtualisation hosts
1 GPU workstation
16 virtual machines running
17 defined

Three virtualisation hosts and one GPU workstation. Sixteen virtual machines running, seventeen defined, one job each: news, actors, situation picture, knowledge graph, agent, git, vault, dashboard, workbench. A guest that needs a graphics card asks the GPU machine for a permit rather than owning one.

Backup

  1. Late AugustThe graph chain dead for four days on all three tiers.
  2. 1 September 2026Found, repaired, and a restore tested.

Two chains. A nightly pull of eight hosts onto one machine, and a separate three tier chain for the knowledge graph and its vectors. It is worth saying what that is worth: in late August the graph chain was dead for four days on all three tiers while the freshness check stayed green. It was found, repaired, and a restore was tested on 1 September 2026. A backup nobody has restored from is a belief, not a backup. The full account is an exploration of its own.

Git at home

On the lab's own git server around seventy
Public on GitHub six
  1. A push.
  2. A pull on the target.
  3. A check against the running service. Because a green build is not an arrival.

Around seventy repositories on the lab's own git server. The public code on GitHub is six of them. Every deploy is a push, a pull on the target, and a check against the running service, because a green build is not an arrival.

The Atlas

  1. One set of documents. Every machine, service, port and interface.
  2. Versioned, linted on every commit.
  3. Distributed to all twelve hosts. An agent arriving anywhere reads the same map.
  4. A copy in the knowledge graph. So it can be searched rather than only read.

One set of documents describes every machine, service, port and interface. It is versioned, linted on every commit, and distributes itself to all twelve hosts, so an agent arriving anywhere reads the same map. The knowledge graph holds a copy so it can be searched rather than only read.

What the Atlas keeps

Counted 9 September 2026

One documentation tree, kept like code: every entry is a row someone can grep. Two kinds of number below. The first row is measured on the machines: the MCP servers configured on the two machines an agent works from, and their tools, counted by talking to each server. The second row is what the Atlas itself holds. One caveat on the API count: it counts services, one host and port per row, not endpoints. The internal dashboard gateway is a single row, and behind it sit 122 routes.

29MCP servers configured
660tools, counted live on 29 of them
31HTTP APIs in the catalogue
122routes behind one of those APIs
73service ports in the catalogue
91timed jobs in the catalogue
42service sheets
7runbooks
15health targets
61documents
1,261commits since 24.03.2026

Where the two disagree, the machines win. On 03.09.2026 the catalogue listed 35 servers and at least 202 tools; the live count found 660. That gap is a finding about the catalogue, and it is on the list.