kaloyandimitrov73/librechatgraphui

Purpose and Entry Points

LibreChat is a self-hosted, multi-user AI conversation platform: unified provider chat, configurable agents, MCP tools, files/RAG, artifacts, and enterprise access controls. The browser application starts at client/src/main.jsx, mounts routing from client/src/routes/index.tsx, and renders the root app in client/src/App.jsx. The Node backend enters through api/index.js and initializes the HTTP server from api/server/index.js.

Frontend Application and State

The primary UI is a React/TypeScript SPA with route-level separation for authenticated chat, startup/login, search, and shared conversations. Chat composition and streamed message rendering center on client/src/components/Chat/ChatView.tsx, client/src/components/Chat/Messages/Content/ContentParts.tsx, and client/src/hooks/SSE/useResumableSSE.ts. API access is isolated behind React-query-oriented data providers, notably client/src/data-provider/connection.ts, client/src/data-provider/Messages/queries.ts, and client/src/data-provider/Agents/queries.ts; local cross-view state is held in Jotai-style stores such as client/src/store/agents.ts.

HTTP API, Authorization, and Persistence

Express-style routes in api/server/routes/index.js dispatch domain APIs for conversations, messages, files, agents, assistants, MCP, administration, and configuration. Middleware establishes request boundaries for JWT authentication, tenancy/resource access, validation, moderation, and rate limits; key paths are api/server/middleware/requireJwtAuth.js, api/server/middleware/accessResources/index.js, and api/server/middleware/limiters/index.js. Database initialization is under api/db/connect.js and models under api/db/models.js.

A newer shared TypeScript domain layer, packages/api, contains reusable business logic and is organized around tenant-aware access, authentication, cache, streaming, MCP, and provider integrations. Persistent document schemas, model definitions, and data-access methods reside in packages/data-schemas/src/schema/index.ts, packages/data-schemas/src/models/index.ts, and packages/data-schemas/src/methods/index.ts.

Generation and Agent Execution Flow

A user message flows from frontend submission through message/agent routes to endpoint-specific clients, then streams incremental events back to the browser. Legacy endpoint assembly is in api/server/services/Endpoints/agents/build.js; agent request and response orchestration is in api/server/controllers/agents/request.js and api/server/controllers/agents/responses.js. The shared agent runtime provides loading, context/memory construction, tool validation, run lifecycle, subagent chains, and human-in-the-loop checkpoints in packages/api/src/agents/run.ts, packages/api/src/agents/context.ts, and packages/api/src/agents/hitl/runtime.ts.

Provider abstraction supports OpenAI-compatible APIs plus OpenAI, Anthropic, Google, Bedrock, Azure, and custom endpoints. Implementations/configuration are concentrated in packages/api/src/endpoints/index.ts, packages/api/src/endpoints/openai/initialize.ts, and packages/api/src/endpoints/anthropic/llm.ts.

Streaming and Distributed Runtime

Resumable generation is a distinct subsystem: jobs, event transport, aborts, approvals, and steering are managed in packages/api/src/stream/GenerationJobManager.ts, packages/api/src/stream/SteeringLifecycle.ts, and packages/api/src/stream/implementations/RedisEventTransport.ts. It can operate in memory or use Redis for multi-instance replay and synchronization. Client handlers consume those events via client/src/hooks/SSE/useEventHandlers.ts and update chat state/rendering.

Tools, MCP, Files, and Retrieval

Tools are normalized and authorized before execution through packages/api/src/tools/definitions.ts, packages/api/src/tools/registry/index.ts, and legacy structured implementations in api/app/clients/tools/structured/. MCP connections add externally hosted tool servers, OAuth, server registries, and SSRF-hardened requests; core boundaries are packages/api/src/mcp/MCPManager.ts, packages/api/src/mcp/MCPConnectionFactory.ts, and packages/api/src/mcp/oauth/hardenedFetch.ts.

File upload and processing are independent from chat rendering: validation, parsing/encoding, RAG preparation, retention, and storage adapters live in packages/api/src/files/index.ts, packages/api/src/files/encode/processAttachments.spec.ts, and api/server/services/Files/strategies.js. Storage supports local and cloud-backed paths, including S3/CloudFront and Azure/Firebase adapters.

Deployment, Observability, and External Services

Container deployment is defined by docker-compose.yml, Dockerfile, and Kubernetes Helm templates in helm/librechat/templates/deployment.yaml. Redis is material for caches, rate limiting, leader election, MCP registries, and distributed streaming; its configuration appears in redis-config/. Optional telemetry routes through Langfuse/OpenTelemetry support in packages/api/src/langfuse/index.ts and the Go fanout service at otel/langfuse-fanout/cmd/langfuse-fanout/main.go. External dependencies include LLM providers, MCP servers, OAuth/LDAP identity providers, object storage/CDNs, search/retrieval services, and optional RAG API deployment charts.