Yarlis CorpusOS

Universal data preparation platform for AI agents. Open source (Apache-2.0), runs on macOS, Windows and Linux, and is not tied to any cloud.

CorpusOS takes the files you already have (PDFs, Word, Excel, slides, email, HTML, code, archives, audio) and turns them into trusted, searchable, permission-aware, traceable knowledge that any authorized LLM or agent can use.

It produces	For	You use it through
A knowledge corpus	answers built from cited, access-checked passages	search in the web app, /api/search, MCP search_corpus
Typed tables	exact figures, with units and formulas kept	the Tables tab, /api/sql, MCP query_table
Platform exports	ChatGPT, Claude, Copilot Studio, Bedrock and Vertex AI knowledge	corpusos export knowledge
Model datasets	evaluation and fine-tuning	corpusos export training, corpusos eval

What happens to a file

discover → store raw bytes (immutable, SHA-256) → detect the type from the bytes → parse
→ PII scan → screen for injected instructions and hidden text → classify → quality gates
→ canonical document → typed tables → [human review if high risk] → chunks → embeddings → index

• A file that fails a quality gate goes to quarantine, with the reason.
• A risky file (restricted data such as SSNs, instruction-like text, hidden text) is stored but held for review. It is not searchable until an admin approves it.
• A changed file gets a new revision, and the old one is kept. A deleted file leaves search, SQL and exports on the next sync.

Ways to use it

• Web app: sign in, search, browse, query tables, upload files and approve held documents.
• HTTP API: the same features for scripts and services, with API keys.
• Storage: put files in a folder or bucket, and they are ingested on the next sync.
• Agents: an MCP server (stdio, or remote with OAuth) and upload-ready exports.
• CLI: corpusos ingest, search, sql, review, scorecard and more. See Getting started.

Every interface checks the caller against the live access list on every call, and audits each call.