Give every AIyour company'scontext.
Connect Notion, Drive, and email once. OpsBlox cleans, structures, and governs your firm’s knowledge — then serves it to Claude, ChatGPT, or any MCP-capable AI, with citations and permissions built in. Built in South Africa for consultancies, agencies, and advisory firms.
§ 01 — The problem
Three reasons
AI with amnesia
Your team already pays for Claude and ChatGPT. Every chat starts from zero — your clients, projects, and history aren't in the room.
Copy-paste context
The real context lives in Notion, Drive, and inboxes. People paste fragments into prompts — stale, partial, and gone by the next session.
Nobody's watching the answers
Sensitive documents end up in prompts, nothing is cited, and there's no record of what the AI saw. Useful, until it's a liability.
§ 02 — The layer
OpsBlox sits between your company’s sources and the AI tools your team already uses. Three jobs: structure the knowledge, map the business, and govern what every AI gets to see.
Knowledge
sync → clean → classify → embed
Every document in the company, cleaned and structured once.
- Syncs Notion, Drive, and email on a schedule
- Cleans, classifies, and summarises every document
- Sensitivity and freshness scored on every record
- Semantic search across everything, instantly
Connects to
Graph + Memory
entities · relationships · evidence
Your business as entities and relationships — not a pile of files.
- Extracts people, clients, projects, and products
- Maps how they relate, with evidence attached
- Proposes memories from real conversations
- A human approves every memory — no silent writes
Extracts
Interaction layer
12 tools · intent → resolve → retrieve → pack
Tools that work the data over MCP — not a passive pipe to it.
- Retrieval router: graph + vectors + memory in one pass
- Entity resolution — “WestCape” becomes the client record
- Context packs assembled server-side, in one call
- Permission-checked, cited, audit-logged — every call
Serves
§ 03 — Deployment
Same layer, same governance, same MCP endpoint — pick where it lives. Most teams start in the cloud and move to on-prem when compliance asks for it.
Cloud Portal
Sign up with your work email. Connect your first source and plug in Claude the same day.
- Isolated tenancy — your own Postgres + pgvector partition
- Hosted in South Africa (AWS Cape Town, af-south-1)
- 15-min incremental syncs · MCP clients from day one
- Encrypted at rest · monthly subscription, scale up or down
On-prem Box
The same layer, on a box in your office. Your documents never leave the building.
- Mac mini installed on-site, managed remotely by us
- Air-gapped option for regulated workflows
- Local models or Claude API — your call
- We handle install, updates, and connector tuning
Same admin, same connectors, same MCP endpoint — moving from cloud to on-prem is a switch, not a rebuild.
§ 04 — In action
One layer, every client
Ask Claude. It asks OpsBlox.
Your consultant asks Claude about a client. Claude calls the OpsBlox MCP server, retrieval is permission-checked, and the answer comes back grounded in your real documents — with sources, confidence, and an entry in the audit trail.
Add OpsBlox to Claude, ChatGPT, or your IDE. One MCP config — no plugin per tool.
Plain questions in the AI you already use. OpsBlox resolves entities, searches knowledge, graph, and memory.
Every result carries its source, confidence, and sensitivity. Restricted documents never leave the layer.
Cloud or on-prem — the MCP surface is identical.
§ 05 — The toolset
MCP is the protocol; the product is what we built inside it. Every question your AI asks is routed — entity resolution first, then graph, vectors, and memory in parallel, assembled into a context pack. These are the tools doing that work.
intent → resolve entity → graph + vector + memory → context pack
Semantic search across every document. Returns chunks with source, confidence, and sensitivity — never an unattributed answer.
The headline act: one call assembles documents, graph context, and approved memory on a topic into a single structured pack.
The short version of any document, with a pointer back to the full source.
Turns “WestCape” into the canonical client record — so every other call starts from the right entity.
Everything connected to a person, client, project, or product: documents, relationships, and the evidence behind them.
How-we-do-things, served as structured steps — your firm's methods, not the model's improvisation.
Approved memory only — semantic, episodic, and procedural. What the company has agreed to remember.
The AI proposes something worth remembering. It lands in the approval queue — it never writes memory directly.
The human gate, exposed as tools — review what the AI wants to keep, accept or decline with one call.
Everything awaiting review, in one queue.
What's connected, how many documents, and when each source last synced.
Not a tool you call — a record you get. Every retrieval logs who asked, what was read, and what was returned.
Structured JSON in and out — your AI writes the prose, the layer guarantees what it’s based on. Custom tools for your firm’s workflows are an Enterprise conversation.
§ 06 — In your week
Nine things your team can do once its AI has real company context — in Claude, ChatGPT, or any MCP client, all cited back to source.
Knowledge
sync → clean → classify → embed
Your AI answers from your actual documents.
Ask Claude for the full history on any client — pulled from Notion, Drive, and email, with sources.
Proposals drafted from your actual past work and pricing — not the model's imagination.
“Where's the latest scope for X?” — answered with the document, not a guess.
Graph
entities · relationships · evidence
Your AI knows how the business connects.
Who connects to what — people, clients, projects, products — with the evidence attached.
See how a supplier touches three projects before you walk into the renegotiation.
One call assembles everything an AI needs on a topic: documents, entities, memory.
Governance
approval queue · filters · audit log
Your AI is on the record, every time.
The AI proposes what to remember; a human approves it. Nothing writes itself.
Restricted documents are excluded from retrieval — by classification, not by hope.
Every retrieval logged: who asked, what was read, what was returned.
§ 07 — How it works
Connect your sources
Point OpsBlox at Notion, Drive, and email — or upload documents directly. The first sync backfills everything, then stays incremental.
oauth connect · full backfill · 15-min incremental sync
We structure everything
Every document is classified, summarised, chunked, and embedded. Entities and relationships are extracted into your company graph, with evidence linked back to source.
classify → summarise → chunk → embed · postgres + pgvector
The interaction layer works the data
Not a passive pipe. Twelve purpose-built tools resolve entities, route each question across graph, vectors, and memory, and assemble context packs — permission-checked, with citations.
intent → resolve entity → graph + vector + memory → context pack
Plug in your AI
Add one MCP config to Claude, ChatGPT, or your IDE. Every teammate's AI now answers from the same governed context — structured data in, cited answers out.
mcp://opsblox · structured json + citations · audit-logged
§ 08 — Pricing
One subscription for the whole company — priced by sources and documents, not seats. Bring it on-prem when compliance asks for it.
[ Cloud Portal · monthly ]
Starter
Solo founders & small teams getting started
- 2 source connectors (Notion + Drive)
- Up to 1,000 documents
- Full MCP toolset · 3 connected AI clients
- Citations + audit log included
Growth
Most growing SMEs · 10–30 staff
- All connectors, incl. email
- Up to 10,000 documents
- Unlimited AI clients across the team
- Knowledge graph + memory approval queue
- Priority email + chat support
Full Stack
Companies running everything through it
- Unlimited documents and sources
- 1 custom connector slot
- Priority sync (sources fresh within 15 min)
- SLA on retrieval latency
- Dedicated account contact
[ On-prem Box · Enterprise ]
From R30,000 setup · R10,000 / month
Mac mini installed on-site, your documents never leave the building. Includes hardware, install, connector setup, and ongoing tuning. 12-month minimum on the retainer.
§ 09 — For consultancies
Already advising SMEs?Bring OpsBlox to your clients.
You're the most credible referrer your clients will ever meet. Run the context layer on your own practice. Introduce it where it fits. We pay 10% on year-1 revenue.
Become a referral partnerA 3-line intro. That's it.
Discovery, demo, proposal, install, ongoing — all ours.
10% on year-1 revenue. Cloud Growth ≈ R1,200 · Full Stack ≈ R2,400 · On-prem Box ≈ R15,000+.
§ 10 — Questions
What people ask before they buy.
What is OpsBlox actually doing day to day?
It's the context layer between your company's knowledge and the AI tools your team already uses. It syncs your sources (Notion, Drive, email, uploads), cleans and classifies every document, builds a graph of your entities and relationships, and serves all of it to Claude, ChatGPT, or any MCP client — with citations, permissions, and an audit log on every answer.
Which AI tools does it work with?
Anything that speaks MCP (Model Context Protocol — the open standard for connecting AI tools to data). Claude Desktop and Claude Code today; ChatGPT, Cursor, and other clients as they ship MCP support. Connecting a tool is one config entry — there's no plugin to build per tool.
Where does my company data go in the cloud version?
Into your isolated tenancy, hosted in South Africa (AWS Cape Town). Each org has its own data partition — not shared across customers, encrypted at rest, and only your authenticated team's AI clients can query it. LLM processing runs on Anthropic's Claude API (no training on your data). For regulated workloads, the on-prem Box keeps everything inside your network.
What does “governed” actually mean here?
Four concrete things. Every document carries a sensitivity classification, and restricted ones are excluded from retrieval. Every answer carries citations back to its sources, with confidence. AI-proposed memories sit in an approval queue until a human accepts them — the AI never writes its own memory. And every retrieval is logged: who asked, what was read, what was returned.
What's under the hood, technically?
A TypeScript service in front of Postgres. pgvector handles semantic search (1536-dimension embeddings); entities and relationships live in lightweight graph tables — no separate graph database to operate. Claude does the document enrichment (classification, summaries, extraction). On top sits the interaction layer: an MCP server exposing twelve tools, with a retrieval router that resolves entities first, then queries graph, vectors, and memory in parallel and assembles a context pack. Everything in and out is structured JSON with citations — your AI writes the prose.
How is this different from uploading files to ChatGPT?
A file upload gives one person, one session, partial context. OpsBlox is shared infrastructure: one cleaned, structured, permission-aware layer that every AI client in the company reads from — plus a knowledge graph and an approved memory that compound over time. Upload-and-chat can't do any of that, and it can't tell you what the AI saw.
Can I move from cloud to on-prem later?
Yes. Same admin, same connectors, same MCP endpoint — switching is a deployment change, not a rebuild. We migrate your documents, graph, and approved memory across.
How long until it's useful?
The first sync runs a full backfill of your sources — a few hours for a typical workspace. After that it stays incremental, re-syncing every 15 minutes. Your team's AI is answering from real company context the same day you connect.
What happens if I cancel?
Cloud: month-to-month, cancel anytime; we export your documents, graph, and audit trail before closing the tenancy. On-prem: the box and everything on it stay yours — we just turn off remote management. 12-month minimum on the on-prem retainer to make the install economics work.
§ 11 — Get started
Start free. No deck.
Drop your work email. We'll provision your tenancy, connect your first source, and hand you an MCP config for Claude. Your AI knows your business by the end of the day.