Mindbase
Mindbase reads the raw record of how a company operates, its chats, tickets, incidents, and pull requests, and turns the decisions people keep making into versioned skills that AI agents can run over MCP. Every skill carries verified citations and is approved by a human before an agent can use it.
Every model knows everything about the world and nothing about your company. Mindbase fixes that. It reads the raw record of how a team actually works, the Slack threads, tickets, incidents, and pull requests, and pulls out the decisions people keep making. Each one becomes a skill, a small piece of versioned decision logic with the exact sources quoted underneath it. A person approves the skill, and then any AI agent can call it over MCP and do the work the way this company does it.
The two halves
Mindbase has two sides. The Cortex runs on the server and holds the company brain. It ingests the sources, extracts the skills, serves them to agents, and watches production. The Neuron runs on each person's Mac and holds their private half. It captures their own AI coding sessions on the device, and only what they choose to promote ever reaches the company brain. The sync cannot carry a transcript, by design.
How it helps a company
Agents that act, not just answer
An agent can run the actual refund or triage the actual incident, following the company's own rules, because the logic is executable rather than a page someone has to read first.
New hires stop guessing
Ask how do we do this and the answer comes back with the exact threads it came from, so people stop interrupting the one person who still remembers.
Incidents resolve faster
When a production signal fires, Sentinel matches it to past incidents, finds the runbook, names the owner, and writes a cited diagnosis.
Knowledge stops leaving with people
Skills and the people graph are built from evidence, so when someone quits, what they knew stays behind and stays queryable.
Trust is earned, not assumed
Every skill carries verified citations and a confidence number that only grows from real outcomes, and nothing runs on its own until it has proven itself.
Skills are the core idea
A skill is decision logic with receipts. It has an owner, a confidence score earned from evidence and capped at 0.90, and the exact sources quoted verbatim under the trigger. If a quote cannot be found in the cited source, the pipeline drops it. No verified citation, no skill.
It knows who knows what
The people graph infers who understands each part of the company from real artifacts, not from self reporting. There are no productivity scores and no failure rankings anywhere. That is enforced by tests, because the moment a tool like this starts ranking people, everyone stops feeding it.
It watches production
Sentinel is the part that earns its keep on a bad day. A signal comes in, and it correlates the deploy, matches past incidents, finds the runbook, and names the owner, then writes a diagnosis with citations. A sandboxed agent can go one step further, reproduce the bug, write a patch, and open a pull request, but it never pushes on its own.
The Neuron keeps people private
The Neuron is the reason people trust it enough to plug in. It captures your AI coding sessions on your own machine, keeps them in a private local brain you can search, and lets you promote only the pieces you want the company to see. The sync payload structurally cannot carry a transcript, and the server rejects anything transcript shaped. Privacy here is the architecture, not a promise.
How it is built
The web app and its API, a deny by default auth layer, and every product surface from the dashboard to the war room.
One boring, reliable store for artifacts, skills, and the vector search behind extraction and retrieval.
Ingests sources into normalized artifacts, then clusters and distils them into draft skills with mechanically verified citations.
The one interface any agent uses to query the approved skills and act on them.
The Neuron, a small menu bar app that captures on the device and syncs only what a person promotes.
What is real today. The web app and the demo are hosted, and the demo genuinely runs. The Neuron, the MCP server, the pipeline, and the sandboxed fix agent are real but local by nature. Still pending, and labelled that way in the product, are live connectors to Slack and GitHub, container grade sandbox isolation, and signal adapters wired to real monitoring. The demo company, Skysail, is synthetic. The data is fake and the processing is real. There are no invented logos, metrics, or testimonials anywhere.
The one line version
Mindbase is the layer between a company's messy history and an AI agent that can act on it. Retrieval ends at an answer. Mindbase ends at a finished task.