Skip to main content

​
What is Cruq AI

Cruq AI lets you build LangGraph agents on your own model provider connections, give them tools, ground them in your own knowledge, and run them with full observability. Everything is tenant-scoped and isolated at the database level.

Agents

Visual, data-driven LangGraph agents with a ReAct tool loop.

Long-running agents

Runs that pause, wait, plan, and check their own work.

Voice agents

Agents that answer and make phone calls in real time, in Indian languages.

Models & providers

Bring your own provider key; address models as provider/model.

Knowledge (RAG)

Knowledge bases with self-hosted embeddings and pgvector retrieval.

Knowledge graphs (GraphRAG)

Graph-structured retrieval over entities and their relationships.

Memory

Long-term facts recalled semantically across threads.

Integrations & channels

Connect services as tools; trigger agents from Slack or Telegram.

Skills & schedules

Procedures loaded on demand, and recurring unattended runs.

Guardrails & Prompts

Input/output safety policies and a versioned prompt registry.

Evals

Compare models on test cases suggested from real runs, and check the judge against your own scores.

Observability

Every tool call and model call on every run, with cost and latency, exportable to OpenTelemetry.

​
Surfaces

  • Web app (app.cruq.ai) - build agents, connect providers and integrations, manage knowledge/memory/skills/guardrails/prompts, watch runs, and compare models.
  • /v1 REST API - programmatic access, authenticated with an API key.
  • cruqai CLI - run agents, stream traces, compare models, and manage resources from your terminal.
  • MCP server (mcp.cruq.ai) - run, inspect and build agents from Claude, ChatGPT, Cursor, Codex or any MCP client, signed in with your Cruq AI account.
See What’s new for recent changes.

​
How it fits together

An agent references tools and a model (via a provider connection). At run time the runtime streams a tool-call trace you can observe live or replay later. Agents can retrieve from knowledge bases and read/write memory through built-in tools, all embedded by a self-hosted embedder so no external embeddings key is required. Every run - interactive, scheduled, triggered from a channel, or part of an eval
  • is persisted with its trace, token usage, cost, and duration. That is what makes a model comparison possible: the numbers on the leaderboard come from the same metering as production traffic.