five layers.
one operating system.
memory, skills, tools, agents, routines — each layer independently valuable, each shaped by the layer below. start with memory. add the rest as you're ready. trust runs through everything.
five layers, top to bottom.
what each layer unlocks.
each layer ships with a first-day action. you don't wait six months to see value — every layer earns its rent in week one.
a single, shared memory of your business that sits between your team and every AI tool you use. ingests slack, notion, drive, gmail, hubspot, salesforce, meeting transcripts, voice notes, contracts and decks. extracts entities, relationships, decisions. writes back from every good answer. compounds the longer you use it. this is the foundation every other layer stacks onto.
skills are the company's playbooks made executable. how you write a deal review. how you triage a support ticket. how you draft a board update. each skill is a versioned unit, governed, and auto-derived from the memory layer plus a 6-week codify sprint — not authored from scratch in a prompt window.
a single, governed gateway to every system your company runs — slack, hubspot, salesforce, jira, github, notion, drive, your bespoke ERP. supports the open MCP standard plus native connectors. read AND write, with per-action approval policies and a full audit trail. this is the moat — the wiring is the work.
agents are typed: sales-development, finance-controller, recruiter, support-tier-1, executive-assistant. each one = a slice of memory + a curated skill set + the tools it's allowed to use + guardrails + a directly responsible human. agents are co-workers, not assistants — they own outcomes.
routines orchestrate agents on a schedule or in response to events. monday 6am: pipeline pulled, weekly review drafted, posted to #revops with the GM tagged. ticket marked p1: support-tier-1 agent triages, reproduces, files the bug — engineering wakes up to a complete brief. this is the autonomous-company piece.
sovereignty (swiss-hosted, on-prem, air-gapped, bring your own AI model). evaluation harness on every skill, agent, routine. full audit log. role-based access, kill-switches per agent. built-in, not bolted on. our default-deny posture is what makes regulated industries say yes.
multimodal capabilities.
multimodal inputs are handled natively. text, images, video, audio. contracts, decks, scanned documents, and meeting recordings become queryable content.
documents, notes, messages
contracts, memos, slack threads, notion pages — all parsed, linked, and cited.
screenshots, diagrams, scans
scanned documents, whiteboard photos, and product screenshots become queryable.
recordings & demos
recorded calls and product walkthroughs get transcribed, summarised, and indexed.
voice notes & dictation
voice memos and podcast clips flow into the same knowledge layer as text.
contracts, decks, reports
layout-aware parsing pulls structure from pdfs — tables, clauses, figures.
transcripts with speakers
who said what, when. decisions and commitments surfaced and linked to the account.
every layer is open.
markdown repo as source of truth. postgres with vector search for retrieval. open standards for the interface. every layer is open. no proprietary lock-in. you can take the knowledge layer with you if you leave.
os & architecture.
do i have to adopt all five layers?
no. start with memory (l1) and stop there if you want — that's a complete product. add skills, tools, agents, routines as your team gets comfortable. each layer is independently valuable; each makes the layer below it more useful.
how is this different from a rag setup or an internal wiki?
rag re-derives knowledge on every query. a wiki sits until someone updates it. blankcollar compiles once, keeps current automatically, builds cross-references, codifies your skills as versioned units, and lets agents act on systems through one governed gateway. memory is a feature, not the whole product.
what about a custom GPT, claude project, or one-off agent we already built?
those are point solutions in one tool. blankcollar is shared across every AI tool, with codified skills (l2), governed tool access (l3), and a trust plane (l0). pilots that fail tend to fail at exactly those layers — see the MIT 2025 GenAI-Divide stat on the homepage.
what does the underlying architecture look like?
markdown repo as source of truth, postgres + pgvector for retrieval, MCP for tools, an evaluation harness on every skill/agent. every layer is open and portable. you can take the OS with you if you leave.
can it read pdfs, images, and meeting videos?
yes. multimodal inputs are handled natively at l1. text, images, video, audio. contracts, decks, scanned documents, meeting recordings.
who this is for.
the product is the same. the focus differs by who's buying and what they need on day one.
install blankcollar.
one layer at a time.
l1 in five minutes via /start. l2–l5 via the kickoff (90 min) or codify sprint (6 weeks). full transformation programme for enterprise — 12 months, co-led, outcome-priced.