CortexGraph · Company Brain Platform

Acompanybrainforeveryagent.

One workspace where you, your custom agents, and every chatbot you already use — Cursor, Claude, ChatGPT, Codex, Gemini — share the same living map of how your company actually works. Ingest knowledge once, expose it over MCP, and let long-running work survive every handoff.

cg_company_brain · agents online
Live
Context graph edges
_
One MCP endpoint · your agents and theirs, on the same brain
  • Cursor
  • ChatGPT
  • Claude
  • Codex
  • Gemini
  • Devin
  • Windsurf
The Bottleneck

Models got smart. Your company didn’t catch up.

How refunds get handled. Why the pricing exception was made. What the on-call engineer did at 2am. The know-how is scattered across people’s heads, Slack, tickets, repos, and old email threads. Humans muddle through. Agents stall, hallucinate, or quietly do the wrong thing.

The Missing Primitive

A company brain agents can actually use.

  • 01One company brain. Every chatbot, custom agent, and human reads from it.
  • 02Knowledge is ingested once and stays current — not re-pasted into every prompt.
  • 03Primitives encode how your business actually decides, escalates, and ships.
  • 04Long-running work survives every model swap, agent crash, and weekend.
The Difference

Same agent. A company brain behind it.

Models are not the bottleneck anymore — domain knowledge is. Without a company brain, your agents reason from a thin slice of context and ship partial work. Connected to one, they plan across every surface your business actually runs on, link their work to durable decisions, and hand off cleanly to the next agent or human.

Agent · no company brain
solo
  1. Read README.md
  2. I'll add the new pricing tier as a record in the pricing module and update the docs.
  3. Edit src/pricing/tiers.ts — added "scale" tier
  4. Edit docs/pricing.md — added tier description
  5. Done. New 'Scale' tier added with a sensible default price. You may want to wire it up to billing.
What the agent missed
  • Stripe product / price catalog — never updated, checkout still 404s on the new tier
  • Marketing pricing page — copy unchanged, contradicts the new tier
  • Sales playbook — discount matrix has no row for Scale, AEs will improvise
  • Customer success runbook — no upgrade path, support will guess
  • Decision log — why Scale exists and what it replaces is nowhere
  • Past Scale-tier discussions in Slack and Linear — never connected
Agent + Company Brain (MCP)
cg.mcp
  1. cg.graph "pricing tier" across workspace17 sources · 6 primitives · 2 active threads
  2. cg.read primitives/decisions/2026-q1-pricing-rewrite.md
  3. This crosses pricing code, Stripe catalog, marketing copy, sales enablement, support runbooks, and the active "Scale tier" thread. Drafting a plan.
  4. cg.plan 7 steps · Cortex routing pricing-eng
  5. Edit 9 files across 5 surfaces · 1 new decision primitive · 1 thread checkpoint
  6. Complete. All surfaces consistent. Decision logged with provenance. "Scale tier" thread resumed for review.
Nothing missed
  • Stripe catalog mutation queued behind human approval
  • Marketing pricing page copy updated and cross-linked
  • Sales discount matrix regenerated from primitives
  • Support runbook updated with new upgrade path
  • Decision primitive captures rationale and supersedes prior tier
  • Thread resumed with checkpoint, follow-up agents notified
The Platform

Ingest. Coordinate. Expose.

Three layers, one substrate. Pull fragmented company knowledge into a maintained brain, coordinate the agents and humans working against it, and expose the whole thing over MCP so every chatbot you already pay for becomes an informed teammate.

01Knowledge layer

Ingest. how your company actually works.

Pull in docs, repos, Slack, tickets, runbooks, and CRM as immutable raw sources. CortexGraph compiles them into a living wiki and typed primitives — with provenance back to every original line, so agents always know where a fact came from.

Raw + Wiki + Primitives

Sources stay immutable. The compiler maintains a cross-linked wiki and typed records — the executable skills your agents read from.

Learn more
Beta

Connectors

Scoped app connections, inbound webhooks, and outbound event delivery. Knowledge gets pulled out of every silo and given provenance.

Learn more

Living Docs

Pages stay current as new sources land. Stale claims flagged. Contradictions surfaced before an agent acts on them.

Built in
02Collaboration layer

Coordinate. your agents and theirs.

Threads, runs, checkpoints, and clarification loops are first-class. Cortex — a real durable agent — notices ambiguity, asks the right human over the right channel, captures the answer back into the brain, and resumes the work. Long-running tasks stop dying in lost transcripts.

Threads & Checkpoints

Multi-day, cross-agent work survives every model swap and handoff. Any agent picks up from the last checkpoint.

Built in

Cortex agent

A coordination concierge that lives in the brain — escalates only what needs a human, then routes the answer back to the right specialist.

Built in

Routing

Clarifications and follow-ups land with the right actor on the channel they actually use. No broadcast. No lost context.

Built in
03Skills layer

Expose. an executable brain over MCP.

Every primitive, decision, and thread becomes machine-readable over a single MCP endpoint. Any MCP-aware chatbot or agent — yours or off-the-shelf — connects in seconds, picks up the same context, and writes back into the brain so the next run starts smarter.

Single MCP endpoint

One token-protected URL exposes threads, questions, checkpoints, triggers, and decisions to any MCP-aware agent in seconds.

Learn more

Executable skills

Decision primitives, runbooks, and patterns become tools agents can call — not paragraphs they have to interpret.

Built in

Activity & runs

Every workflow leaves an inspectable trail. Resumable, idempotent, and queryable — including by other agents.

Built in
“Cursor, Claude, and our internal agents are now reading from the same brain. They stopped re-asking the same questions and started handing work off to each other — and to us — without dropping context. It’s the missing layer between our company and our agents.”
Field note · early-access AI-native team

From scattered knowledge to
an agent-ready brain.

01

Spin up your company brain

One shared workspace per company — humans, custom agents, and chatbots all reading from the same map.

02

Pull knowledge out of the silos

Connect docs, repos, Slack, tickets, CRM, and runbooks. Sources stay immutable; the brain stays current.

03

Define your primitives

Encode how your business actually decides and ships — accounts, decisions, runbooks, escalation paths.

04

Plug in every agent over MCP

Cursor, Claude, ChatGPT, Codex, Devin, your in-house agents — one OAuth and they all start informed.

05

Run long-horizon work

Agents collaborate across days and tools. Cortex resolves ambiguity. Checkpoints make every run resumable.

Coordination Layer

Long-running agents
that ask before they guess.

When any agent — yours or one of the chatbots you already use — hits ambiguity, conflicting context, or a decision it shouldn’t make alone, Cortex steps in. It pings the right teammate on the channel they actually read, captures the answer back into the company brain, and resumes the work. No silent stalls. No hallucinated policy.

CortexiMessage
Agent 'eng-planner' is drafting the Q3 roadmap and hit a fork. The strategy doc references a "Phase 2 pivot" but no decision primitive exists for it.
Should I treat the pivot as approved, and what changed about the scope?
Approved last week. We dropped the standalone app and are shipping it as an MCP server instead.
Logged as a decision primitive, linked to the strategy doc, and resumed 'eng-planner'.
iMessage
Built for AI-native scale

Trusted. Inspectable. Yours to run.

A company brain holds the most sensitive thing you own — how your business actually works. Every layer is auditable, every agent action is provenance-linked, and the substrate is yours to host where your data already lives.

  • SOC 2-ready by design

    Markdown stays canonical. Postgres owns provenance, runs, and activity. Every mutation is auditable from the first commit.

  • No vector lock-in

    Search and retrieval work without embeddings. Bring your own ranking models — never required to ship.

  • Clerk-backed identity

    Custom CortexGraph auth screens backed by Clerk sessions. Agent tokens stay separate, scoped, and auditable.

  • Bring your own data plane

    Neon Postgres for indexes, Cloudflare R2 for canonical blobs. Self-host or run on the hosted control plane.

  • Resumable workflows

    Long-horizon work runs on Vercel Workflow — idempotent, inspectable, and resumable across days.

  • Open MCP surface

    Standard Model Context Protocol over OAuth. Any MCP-aware agent connects in seconds. No proprietary client.