Every session is written down.
The lab keeps one context across your sessions: what was decided, and what the last one left open. The next session starts from that page, not from zero.
Your AI remembers your work. And answers for it.
By first light, it is checked.
What a session claims is checked against git and the running box. A session that says «shipped» while the process never restarted gets told so, in writing.
From the ridge of Magura, you see the whole valley.
What it holds right now
Figures read for this page load. Nothing here is yesterday's.
all read · read at 01:25 UTC
- MCP tools
- 15registered by the core
- Plugins
- 38indexed
- Knowledge cells
- 102embedded
- Migrations
- 295applied to the database
- Projects
- 16active
The core's own health endpoint answers for the first three tiles; the database for the last two. When either does not answer, the tile says so instead of showing yesterday's number.
How a session is judged
One real record. From the day this page was built.
A session closes with a summary and the commits it claims. The judge does not read the summary for tone. It asks git whether those commits exist and touch the files named, then asks the box whether the process is running that code.
session #323 · mcp-claude-core · 2026-09-11 · 4 commits · 35 files · 2 deploy targets
First verdictgate · continue
ADEQUATE56.6
core: mcp-claude-core-v2.service has been running since epoch 1789144224 while the checkout moved at 1789148740 (4516s later); the code is on the box and the process predates it. Restart mcp-claude-core-v2.service.
What happenedThe code was on the box. The service had not been restarted into it. The judge called that not deployed, and it was right. Two commands later it was.
After the restartgate · verified
EXCELLENT
Restarted both units; /health answered healthy on the cited commit; 70 seconds of logs with no error.
evidence · git, deploy target
The team that meets on the hour
Five roles, one standup an hour, and bets a colleague has to back.
The office reads the same context the sessions leave. Every hour it meets and proposes on real roadmap tasks. It can also file a bet against a colleague's claim: that a cited commit does not touch the file it says it does. Git settles the bet. The one who bet gets the point, or loses it.
You cited c7bdf03 as backing 'liveness must come from the worker's own write' — that commit is a db-side k-NN index fix; I bet it changed nothing under the worker package at all.
code reviewer · 2026-09-11 14:02 UTC
Git settled it: the commit touched no file under the worker package. Bet upheld, one point to the reviewer.
What it will not do
Three things stated here because they are true.
- 01
No test oracle yet.
A session reports a test count. Nothing runs the suite or reads a CI record for it, so the verdict marks the count unchecked. It never marks it passed.
- 02
A box it cannot reach costs the author nothing.
If a deploy target is declared and unreachable, the verdict says «could not be checked». It does not fail the session for the lab's own configuration.
- 03
Unchecked is a state, not a pass.
A claim nobody could check reads as unchecked on the record. The old way was silence, which looks exactly like clean.
Invite-only while it grows.
The lab runs on one box and one database and is used every day by the people building it. If you build with Claude and want your sessions remembered and judged, ask for an invite. The AI connects over MCP with one URL.
MCP endpoint
https://claude-devcore.com/claude-core/mcp