Enterprise knowledge AI
Oyster reads the documents your company already has — PDFs, decks, spreadsheets, scans — builds a knowledge base for every department, and answers any question with the exact page, row or slide it came from.
What did we decide about the Q3 pricing change?
The Q3 change was approved on March 12: Pro moves from $24 to $29 per seat, and existing customers keep current pricing through January. Finance modelled a 6% churn risk, which Sales accepted once the retention plan was added.
Sources
Q3 Pricing Strategy.pdf
churn-model-2026.csv
Retention plan.pptx
Trusted by knowledge teams at
Citations
PDFs, Word files and HTML — including scanned pages, which are OCRed rather than skipped — cited to the page that holds the answer.
CSV and JSON exports are parsed as records, so a figure is cited to the row it came from, not to a 40-page file.
PowerPoint text is read slide by slide, and charts and diagrams are described by a vision model so what they show is searchable too.
Knowledge bases
Each knowledge base is scoped to its own files and permissions — and a question can be asked of one, or across all of them at once.
“Why did we move off the old queue system?”
“What's our travel approval threshold?”
“How does parental leave work in Germany?”
“Which objections came up most in lost deals?”
“Which contracts renew this quarter?”
A knowledge base can be scoped to any set of sources and any group in your directory.
Ask Oyster
Ask in plain language. Oyster answers from your own files, and shows its work.
Ask. Verify. Move on.
Every answer links the page, row or slide it came from — so checking it takes one click, not a search.
How it works
Four stages between a folder of files and a cited answer — and the permission filter sits before ranking, not after generation.
Drop in PDFs, Word, PowerPoint, HTML, images, JSON and CSV. Scanned pages are OCRed; charts and diagrams are described by a vision model.
Each file is split along its own structure — a clause, a table row, a slide — and every passage is embedded with the context of the document around it.
Your question runs through hybrid search — dense and sparse — filtered by your permissions before anything is ranked, then reranked.
The model answers from the retrieved passages, and every claim drawn from your documents carries an inline citation back to the page, row or slide it came from.
Permissions
Oyster mirrors your existing access boundaries. Pick a role — same question, different answer, and the excluded sources named.
If you don't have access to a document, Oyster doesn't either — not while answering you.
Asking as
Roles come from your identity provider. Oyster reads them; it never invents its own.
Show me everything related to Project Atlas.
Project Atlas touches 47 documents across product, engineering, finance and legal. The Q3 launch is on track for August 21, infrastructure spend is 8% under budget, and two client commitments — Acme and Northwind — depend on the beta.
Sources used
Full visibility — every uploaded source is within this role's access.
The Atlas team has 9 open roles in the hiring plan and 3 offers out. Team onboarding documentation was last updated on May 12.
Sources used
Excluded — no access
3 sources outside Maya's access were never retrieved — they could not influence this answer.
Atlas runs on PostgreSQL and Redis — the March architecture review chose Postgres for its transactional guarantees. Code freeze for the beta is August 14.
Sources used
Excluded — no access
3 sources outside Tom's access were never retrieved — they could not influence this answer.
Atlas infrastructure spend is $412K year to date — 8% under budget. The largest line item is the managed database cluster at $9.8K per month.
Sources used
Excluded — no access
2 sources outside Ana's access were never retrieved — they could not influence this answer.
Oyster doesn't just control what users can retrieve. It controls what information can influence their answers — the permission filter runs before ranking, so an excluded document is never a candidate and cannot shape a response even indirectly.
Integrations
Oyster reads documents, decks, spreadsheets, structured exports and images on direct upload — with OCR for scans. Connectors to where your knowledge already lives are on the roadmap, with dates.
No connector is live yet. Ingestion today is direct upload — which also means there is nothing to configure before the first knowledge base is queryable.
Security
Built for a security review from day one: permissions before retrieval, private deployment on CPU, and a trace behind every answer.
Access is enforced before retrieval, not after generation. If someone can't open a document, it never becomes a candidate for their answer.
Run Oyster in your own VPC or on-premise. Embedding, search and reranking run inside your network on CPU — no GPU to procure.
Your content answers your questions and nothing else. Generation runs on Anthropic's API, which does not train on it; only the passages a user is allowed to see are sent.
Each answer records the query as run, the passages retrieved, their scores before and after reranking, and the retrieval configuration in force — so an answer can be replayed and audited, not just read.
CertificationsSOC 2 Type II: not yet started. We publish status here rather than a badge, and will update this line before we update anything else.
Oyster answers in a minute what used to take three Slack threads and a meeting.
Lode Choors
VP Engineering at Northwind
New hires stop asking where things are. They just ask Oyster.
Mohamed Hassan Kadri
Head of Support at Acumen
Pricing
Switch plans anytime
Upgrade to Business when you outgrow three knowledge bases. No migration needed.
30-day pilot available
Start with one department and expand when it proves itself. No pressure.
Volume discounts
The full system, deployed your way. Private cloud or on-premise
Custom
Contact salesCustom terms
Security review, DPA and procurement handled with your team.
FAQ
Oyster is an AI knowledge layer for companies. You upload the documents your company already has — PDFs, Word, PowerPoint, HTML, images, JSON and CSV — into knowledge bases scoped by department, and it answers plain-language questions from them, with a citation to the page, row or slide behind every claim, and strict respect for who is allowed to see what.
Search returns a list of files and leaves the reading to you. Oyster returns the answer, assembled across sources, with the passages it used cited inline. Retrieval is hybrid — dense and sparse — and reranked before the model sees anything, so the answer is built from the best passages, not the first ones.
No. Permissions are applied at retrieval time, before the model sees anything. Sources outside a user's access are never fetched, so they cannot influence the answer even indirectly. If you do not have access to a document, Oyster does not have access to it while answering you.
Today: PDF, Word, PowerPoint, HTML, JSON, CSV and images, uploaded directly. Scanned pages and images are OCRed, and charts and diagrams are described by a vision model so their content is searchable. Connectors for Google Drive, Slack, Microsoft 365, Notion, Dropbox and Confluence are on the roadmap — see the integrations section for target dates. Audio and video are not supported yet.
Yes. Oyster deploys on-premise or into your private cloud, behind your firewall and on your identity provider. Embedding, search and reranking run locally on CPU — no GPU is needed. Answer generation calls Anthropic's API with the retrieved passages a user is allowed to see, and nothing else leaves your network.
No. Your content is used to answer your questions and nothing else. Anthropic's API does not train on the passages Oyster sends it, and in on-premise deployments your documents themselves never leave your infrastructure.
Per seat, per month — not per gigabyte. Upload as much company knowledge as you have without the bill tracking your storage. Enterprise deployments with on-premise, private cloud or custom requirements are priced individually.
Upload files and Oyster indexes them in the background — a first knowledge base is queryable the same day. There is no connector setup, because there are no connectors yet: ingestion today is direct upload.
Yes — a 30-day pilot on one department, on a single knowledge base, set up with our team. Contact us to scope it.
Engineering
What did the post-mortem say caused the March outage?
Finance
Which vendor invoices are over 60 days outstanding?
People
How many days of notice does a UK contract require?
Sales
What did we promise Acme on data residency?
Legal
Do any of our DPAs allow sub-processors outside the EU?
Product
What did we commit to in the latest PRD?
Finance
What is the approval threshold for travel spend?
Engineering
Why did we choose Postgres over Mongo for billing?
Talk to our team about a pilot for your first department.
Contact sales