Entry — Browser
Users & administrators
Sign in and work in the Unique AI application
Entry — API / SDK
Customer applications
Use the platform headless through the Public API & SDK
Entry — A2A
Other agents
Call Remy from another agent or agent platform
Agent-to-Agent (A2A)Agent as a tool (MCP)
Enterprise Identity Providers
Customer-managed identity
Lives in the customer environment, federates into IAM
Entra IDOkta
Identity & Tenancy
IAM
SSO · SCIM · user & group lifecycle
Organisation Structure
Multiple isolated organisations per instance
Org AOrg BOrg C
authenticated, organisation-scoped requests
Connectors
Tool connectors
MCP Servers
Enterprise · partner · custom
Web
Public internet sources
plug into the MCP Hub
Knowledge sources
SharePoint / OneDrive
Confluence
Google Drive
Salesforce
Teams / Outlook
User uploads
files + access rights feed indexing
Agent: Remy
Conduct
Full agent harness · filesystem · code execution · skills · sub-agents
Security boundary · no secrets inside the sandbox · no direct internet access · strict isolation per user and run
Sandbox filesystem
Skills
Loaded from the Skill Directory
Uploaded files
Brought in by the user
Client Memory & Knowledge Base Files
Injected client memory and other knowledge
Checkpoint
State of the previous run
Artifacts
Files created by the agent
Swappable agent harness
Claude Code
Codex
Cursor CLI
Pi
MITM egress proxy
The only way out of the sandbox. Every outbound request passes through it.
Secret injection
Credentials added here, never seen by the agent
Allow / deny lists
Every destination checked against policy
Request & response filtering
e.g. DLP systems integrated inline
the agent picks its tools
Tools
MCP Hub
connects the MCP servers from the connector column
Virtual MCP Servers
Combines tools from different MCP servers into a single virtual MCP server
Agent to Agent (A2A)
Delegates to other agents via the A2A protocol
Web Search
Built-in internet research
Internal Search
Permission-aware retrieval
Knowledge baseInternal Search queries the knowledge base
Knowledge base — Governed Data Plane
Client Memory
One profile per client of the relationship manager
Gathers what is scattered across CRM, personal notes, meeting notes and the RM's own memory
Kept and transferable, e.g. when a client moves to another RM
Lets Remy tailor services to the client and surface connections otherwise missed
Skill Directory
Reusable skills Remy loads on demand
Versioned skill packages: instructions, scripts, references
Curated per organisation · customer-authored skills
Loaded into the agent sandbox at run time
Virtual File System
Files are stored inside Remy
Mirrors the source folder structure
Access rights copied from the source, e.g. SharePoint permissions
Re-synced on change · every read is permission-checked
Indexingwrite path · asynchronous workers
Ingest & sync
Pull from connectors · delta detection · fetch permissions
Parse & understand
Layout · tables · OCR · image understanding
Chunking
Context-aware · tuned for financial documents: sections, tables, footnotes
Embedding
Dense vectors produced through the Model Gateway
Index write
Sparse and dense indices · metadata and access rights kept per chunk
Retrievalread path · one pass per query, called by Internal Search
Query
Scoped to the user's permissions
Hybrid search
Sparse
Elasticsearch · keyword
Dense
Qdrant · vectors
Fusion
Both result sets merged and ranked with Reciprocal Rank Fusion (RRF)
Return
Chunks with metadata retained, citations point back to the source file
Storage layer
Object Storage
Files & extracted assets
PostgreSQL
Metadata · permissions · state
Qdrant
Dense vector indices
Elasticsearch
Sparse / full-text indices
Operations
Benchmarking
Model and prompt evaluation against reference sets
Audit Logs
User · model · tool activity
Usage & Observability
Quality · latency · cost · adoption · analytics and reporting
PostgreSQL — Remy foundation
Organisations · Spaces · access · chats · tables · MCP configuration · encrypted at rest
Model Gateway & Models
Model Gateway & Policy
Routes every request · model-agnostic · configurable per organisation and Space
Language models
OpenAI
Anthropic
Mistral
Qwen
Llama
···
others
Embedding models
Serves Remy, its tools and its knowledge plane, where it produces embeddings and document understanding.
Customer-owned model endpoints supported · no prompt or completion data leaves the deployment.
Hosting
Everything above runs on one deployment target per customer.
Azure
AWS
Google Cloud
Private cloud
On-premises
No dedicated page in the public documentation yet.