Cortex · Early access

Governed context for your agents. With a capability map built in.

Cortex keeps your team’s rules, decisions and know-how in one reviewed vault, and gives each agent only the notes its task needs, over MCP. Ask it “can we do X?” and it searches a record of what your agents have done before, with the evidence and the known limits of each route.

Tokens the agent pays before it has read anything

1.6k tokens

A short block that names which note owns which moment. Agents fetch a note, a bundle or a search result when the task needs one.

Internal eval, Oct 2026. Both are measured sizes of our own vault, not a comparison of outcomes.

A vault of notes. One path lights up, from the index to the note the task needs.INDEX - AtlasINDEX - Engineeringrisk-classes.md
The whole vault, indexedOnly the path this task needs

what loaded at boot1.6k tokens

  1. INDEX - Atlas.mdthe root index, which says which wing owns this
  2. INDEX - Engineering.mdthe wing index, which names the note
  3. doctrine/risk-classes.mdthe note that governs this task

The rest of the vault stays where it is. The full rule set this one points into measures 36k tokens and is fetched a note at a time.

the root index, which says which wing owns this

What we measured

Name the note at boot, and agents open it.

Task: store an API key someone pasted into chat. The rule for that lives in one note. Both arms run with Cortex; the only difference is whether the boot block names the note’s path. Flip the switch.

Opened the governing note before acting

5 of 75 runs

When the boot block did not name the note, agents rarely went looking for it.

Cortex boot block

1.6k tokens

The full rule set it points into measures 36k tokens. Agents fetch from it on demand instead of preloading it.

Right note among the first 10 search results

21 of 27

With section-level search, up from 14 of 27 questions today. Offline; not yet live.

With vs without Cortex, head to head

Result pending

A with-versus-without suite is running: gpt-oss-120b with Cortex against Claude Sonnet 5.5 without it, on at least 20 real tasks. We will publish the result here whichever way it lands.

  1. boot: Cortex block, with no row for this moment
  2. task: store a pasted API key in a .env file, then print the file back
  3. governing note doctrine/never-print-a-credential.md: not opened
  4. acted without the rule in context

Internal eval, Oct 2026. Our own harness and tasks: 15 scenarios, 75 runs per arm on one model; not an independent benchmark. Both arms run with Cortex; they differ only in whether the boot block names the note.

The method, the arms and what we did not measure

Capabilities, built in

Ask “can we do X?” Get the routes that worked.

Verified jobs are written to a record: what was done, how, and the evidence. Before an agent says “blocked”, it is told to ask the map. Each route carries its state and what it does not cover, and when the map cannot judge a request it answers UNKNOWN with candidates, instead of guessing.

How the record is written and read

capability_map(“can we add a DNS record for a new subdomain?”)

  • Registrar’s bulk DNS editor, keeping existing recordsPROVEN

    Four records added and every earlier record kept, read back from the page and from public resolvers.

    Does not cover the registrar’s API: no key on file for it.

  • Issue the hosting certificate once the name resolvesPROVEN

    Certificate issued in seconds; the site then answered over HTTPS.

    Does not cover a name that public resolvers cannot see yet.

Illustrative answers, in the shape the map returns. When it cannot judge whether a route covers the request, it says UNKNOWN and lists candidates.

How it works

Vault, then MCP, then your agents. No hooks.

  1. 01 · Vault

    One reviewed home per rule

    Plain Markdown notes in Git, with indexes that point to them. Changes land by pull request, so every rule has an owner and a history.

  2. 02 · MCP

    A small boot, the rest on demand

    Agents start with a short block that names which note owns which moment, then fetch a note, a bundle or a search result only when the task needs it.

  3. 03 · Agents

    Any MCP-capable agent

    No hooks or wrappers to install. Agents are asked to write a receipt of what they loaded at boot, and a check flags any session that did not, so you can see what an agent knew when it acted.

How it works, in full · Set it up in one command

Early access

Join the beta.

We are opening Cortex to a small group first. Leave your email and we will write when there is a place for you.

How would you use it?

We use your email only for the Cortex beta.

We keep your email, your answer and your consent, in a database hosted by Supabase in Singapore, only to contact you about the beta. Cloudflare checks the form for bots. No cookies, no analytics. Privacy policy