# RAGfly — Your AI, answering from your documents

> Choose your folders and talk to your documents, with the source of every answer. Connect them to Claude, ChatGPT or your own agents. We don't store your documents.

## HOW IT WORKS

1. Choose your folders. With RAGfly Desktop, on Mac or Windows. RAGfly reads, organizes and indexes them on its own — scans included. We don't store your documents.
2. Try it in the chat. Ask RAGfly's chat about your documents and see how it answers. Every answer comes with its source: the document and the page.
3. Connect your agents. Then put it to work: Claude, ChatGPT or your own agents query your documents whenever they need them. Over MCP, REST, the SDK or the CLI.

## Agent interface catalog

Discover MCP tools with tools/list. Discover the separate manifest operations with GET /v1/operations.

Updated: 2026-10-27

## MCP connection

- Streamable HTTP endpoint (recommended): https://api.ragfly.ai/mcp-http/
- SSE endpoint (legacy, not recommended; its session is lost on a redeploy): https://api.ragfly.ai/mcp/sse
- Authentication: OAuth 2.1 or Authorization: Bearer rf_xxxxxxxxxx (API Key issued from a human session)
- Scope: Every call is filtered by the key owner’s organization, workspace, area and role. Flexible workspace focus is stored server-side.
- RAGfly Desktop is included on every plan, Free included: https://ragfly.ai/download
- Key-based setup script (validates the key, stores it in ~/.ragfly/ragfly-mcp.env, registers Claude Code / Codex / Claude Desktop; download, read, then run): https://ragfly.ai/integradores/ragfly-mcp-setup.sh
- Same setup for Windows PowerShell 5.1+ (not yet run on a Windows machine; see MCP.md): https://ragfly.ai/integradores/ragfly-mcp-setup.ps1

API keys are issued through a signed-in human session. A fixed-workspace key cannot change workspace. A flexible key can select an authorized workspace or release focus with workspace_code: null; the server persists the focus for that key.

## MCP tools (42)

### session

Verifies the connection and returns the identity and the active organization, workspace and area

### set_active_workspace

Sets an authorized workspace for a flexible key or releases focus with null; fixed keys cannot change workspace
Parameters: `workspace_code` (string or null)

### unset_active_workspace

Releases the workspace focus (same as `set_active_workspace` with null): back to the identity's default workspace scope

### list_operations

What this key can do: every operation its RBAC allows, with `kind` and `confirm_required`

### get_operation

One operation with its `input_schema` and `output_schema`
Parameters: `code`

### run_operation

Runs one operation. A `write_confirm` one only runs with `confirm=true`; otherwise it returns a preview
Parameters: `code`, `input?`, `confirm?`

### list_documents

Lists organization documents with filters
Parameters: `status`, `limit`, `page`

### get_document

Full detail of a document
Parameters: `document_code`

### document_edges

Corpus-graph edges of a document: neighbors and documents at 2 hops
Parameters: `document_code`, `neighbor_limit?`

### list_collections

Lists the organization's Collections
Parameters: `limit`

### get_collection

Collection detail: criteria + documents + queue
Parameters: `collection_id`, `doc_limit`

### compose_collections

Set algebra (COMPOSE) of two Collections → a new Collection handle
Parameters: `operation`, `collection_id_a`, `collection_id_b`, `name?`, `collection_type?` (`TEMPORARY`|`SAVED`)

### read_collection

Materialize a Collection (READ) at a chosen resolution, paginated
Parameters: `collection_id`, `resolution?`, `query?`, `limit?`

### refresh_collection

Re-applies the Collection's natural-language criteria and re-materializes its set (picks up newly qualifying documents)
Parameters: `collection_id`

### save_collection

Saves a temporary Collection (`TEMPORARY`) as a saved one (`SAVED`)
Parameters: `collection_id`

### wiki_index

Index of the compiled-knowledge pages visible from the active area. **Not available to integrator keys today (403)**
Parameters: `area_code?`

### wiki_page

One compiled-knowledge page. **Not available to integrator keys today (403)**
Parameters: `document_code`

### compile_collection

Compiles a Collection (background job). The leaf skill must be one your access can run; the default public code is `COMPILE_PAGE`
Parameters: `collection_id`, `leaf_skill_code?`

### queue

Current state of the processing pipeline
Parameters: `process`, `status`, `limit`

### list_runs

Skill run history
Parameters: `limit`

### catalog

User capabilities: available functions + LLM skills (RBAC-filtered)
Parameters: `type?` (`FUNCTIONS`|`SKILLS`|`ALL`)

### list_skills

LLM skills available to you (the same list as `catalog`)

### get_skill

Skill detail: type and output; prompt and model only if your role administers skills
Parameters: `skill_code`

### run_skill

Queues a run over a collection or document
Parameters: `skill_code`, `collection_id?`, `document_code?`

### search_documents

Direct semantic search over the corpus. `filter` narrows by document type, file attributes and content characteristics (codes from `list_document_types` and `list_characteristics`); a valid filter answers `filter_not_available` until structured filters are enabled for the organization
Parameters: `query`, `limit?`, `min_similarity?`, `workspace_code?`, `collection_id?`, `filter?`, `location_code?`

### list_document_types

Document types of the workspace as a tree, with parent, synonyms and document counts. Picking a type also covers its more specific types
Parameters: `workspace_code?`

### list_characteristics

Content characteristics (a date, an amount, an ID, a person…) available for the given types and their more specific ones
Parameters: `document_types?`, `workspace_code?`

### ask

Natural language question with full RAG (non-streaming). `mode: "help"` answers questions about RAGfly itself instead, without links to web screens
Parameters: `message`, `function_code?`, `conversation_id?`, `title?`, `location_code?`, `mode?`

### set_active_area

Set or release an organizational area focus
Parameters: `area_code` (required; `null` releases)

### list_areas

List visible organizational areas
Parameters: `workspace_code?`, `parent_code?`, `query?`, `limit?`, `cursor?`

### list_locations

List visible document folders
Parameters: `workspace_code?`, `parent_code?`, `query?`, `limit?`, `cursor?`

### get_agent_context

Authenticated layered prompt, identity, allowed tools and limits for Agentic Retrieval. Tool names are stable English identifiers; the available list and schemas depend on identity/profile
Parameters: `function_profile?` (`user_chat`|`support_chat`)

### run_agent_tool

Runs one tool from the current authenticated AgentContext
Parameters: `public_name`, `arguments_json?`, `function_profile?`

### get_organization

Reads this organization's profile and what is still missing

### update_organization

Writes the profile. Needs an administrator of the organization
Parameters: `group_description?`, `group_system_prompt?`, `workspace_description?`, `workspace_system_prompt?`

### draft_organization

Proposes the four profile texts from a source text. Does **not** save
Parameters: `source_text`

### get_usage

Plan quotas against what is already consumed

### list_conversations

Conversation history; its ids feed `ask`
Parameters: `function_code?`, `limit?`

### delete_conversation

Deletes a conversation and its messages. Not reversible
Parameters: `conversation_id`

### list_processes

Process instances: support, requests, collection jobs
Parameters: `status?`, `process_type?`, `category?`, `mine?`, `only_open?`, `limit?`, `page?`

### get_process

One process in full, with description and comments
Parameters: `process_code`

### update_process

Updates the fields the "My Processes" screen lets a person edit
Parameters: `process_code`, `status?`, `priority?`, `name?`, `description?`, `comments?`, `assigned_to?`, `due_at?`, `finished_at?`

## Manifest operations

Separate manifest catalog of screen operations allowed by the credential after RBAC filtering. Its entries are not MCP tools.

Endpoint: GET https://api.ragfly.ai/v1/operations

## Guidance resources (1)

### Configurar conectores de ingesta: Google Drive y Dropbox

Cuando una persona necesite conectar Google Drive o Dropbox, acompáñala con estas guías. No es una tool MCP, un comando CLI, una operación REST ni una función del SDK: el administrador configura la aplicación y los parámetros del grupo directamente en RAGfly; cada usuario autoriza su propia cuenta y elige la carpeta en el navegador.

Security rules:
- Nunca pidas, recibas, almacenes ni repitas secretos de cliente, access tokens, refresh tokens, contraseñas, cookies o códigos OAuth.
- Indica al administrador que ingrese las credenciales de aplicación solo en Parámetros de Grupo de RAGfly; el agente solo describe los pasos y verifica requisitos no sensibles.
- No intentes completar la autorización OAuth ni seleccionar carpetas fuera del navegador y la sesión del usuario.

Documentation:
- [Guía humana: Google Drive](https://ragfly.ai/build/google-drive)
- [Guía humana: Dropbox](https://ragfly.ai/build/dropbox)
- [Referencia técnica: Google Drive (Markdown)](https://ragfly.ai/integradores/GOOGLE_DRIVE.md)
- [Referencia técnica: Dropbox (Markdown)](https://ragfly.ai/integradores/DROPBOX.md)