RagLeap
Blog

RagLeap Core: self-hosted AI employees that ask before they act. What it is, what it is not and what is verified

RagLeap Core is an open-source (MIT), self-hosted engine for AI employees that ask before they act. It answers from your own documents, with your own API key, on your own infrastructure. This overview is written from the project README and release notes, and says what is verified and what is not.

What is RagLeap Core?

The README describes a document-grounded chat engine: upload your documents, ask questions, get cited answers. On top of that sit 46 role-based AI employees with an approval gate: webhook, Slack, email, CRM and calendar actions are opt-in through .env and all sit behind the same approval gate. WhatsApp, Telegram and Discord channel adapters are included, single-tenant and configured through .env.

What does it do today?

  • Document ingestion for 28 file types, or a public web page by address, with retrieval over pgvector, hybrid search (dense plus full-text, fused by reciprocal rank fusion) and answers that cite their sources.
  • Bring your own AI key: OpenAI, Gemini, Anthropic or any OpenAI-compatible endpoint, with automatic fallback to a backup provider, streaming answers and real per-call token counts. A local mode runs chat and embeddings through Ollama (small local models on a CPU are slower and less accurate than Gemini, as the README says).
  • Approval-gated autonomy in three modes (off, semi, full). Sensitive roles always need approval, and after the agent reads web or MCP content, outgoing actions need approval even in full mode. Since v0.13.0 that also applies to answers built from your ingested documents.
  • The AI Office dashboard at /office: an approvals inbox, employees by department, tasks, agent runs, activity, settings and a setup checklist. It has no chat screen yet.
  • Opt-in tools: an act-observe agent loop (up to 4 steps by default), sandboxed Python and shell in a network-less container, an MCP client with an allowlist, scheduled check-ins, and optional daily and monthly token caps.
  • Custom roles through PATCH /employees/{role}, a Neo4j knowledge graph beside vector search, language detection across channels, and n8n workflow automation after a reply.

Some integrations (the Meta WhatsApp, Google Calendar and HubSpot items in v0.12.0) were listed from merged pull requests and not separately tested; the roadmap says so.

How do I install it?

The README’s fastest way checks Docker, clones the repo and sets up .env. On Windows, run it in Git Bash, not Command Prompt or PowerShell:

curl -fsSL https://raw.githubusercontent.com/ragleap/ragleap-core/main/install.sh | bash

The manual way, if you want to read the code before running anything. You need Docker, Docker Compose, and either a Gemini API key or a local Ollama setup with no key:

git clone https://github.com/ragleap/ragleap-core.git
cd ragleap-core
cp .env.example .env
# add your Gemini API key to .env
docker compose up --build -d

After installing, the ragleap command starts, stops and updates the stack (ragleap update backs up the database first). To ask a question without writing code, the README points to a small read-only client script or to one of the chat channels.

What is RagLeap Core not?

  • Not a hosted product today. It is self-hosted software you run yourself. Hosting is planned inside ragleap-core, with no date.
  • Not multi-tenant today. It is single-tenant: one document set per deployment, served by a staff of AI employees, not by one bot. Multi-tenant use is planned, with no date.
  • Not a no-code SaaS dashboard: it is a codebase you deploy and configure. Not feature-complete: it is the foundational subset. An embeddable web chat widget is planned and not built.

What has been tested by others?

One community tester ran the Windows installer with the Ollama option on one Windows 11 Pro machine: the services started, the dashboard opened, the connection tests passed, one PDF was ingested and one chat request was answered. That is one machine, one PDF and one question, not a compatibility guarantee; macOS and Linux desktops are not verified. See what was shown and what was not and the evidence.

What is not verified?

  • It is not independently security reviewed. The offline install has not been tried on an air-gapped machine.
  • Single sign-on is off by default and, per the v0.13.0 release notes, was tested on Keycloak for sign-in and refusals only, not on Entra ID or Okta.
  • The Stress Office guardrail test is small: two models, 100 counted injection runs, simple injections, not a security certification.

Where do I go next?

The roadmap lists what is shipped, experimental and planned, and the Vision page the longer-term themes, all labelled Planned. The library has the separate Python, Java and JavaScript libraries, each with its status and limits. The Tests page counts test files and functions for the core app and each library (a count, not a pass count). The latest releases are in the release notes, and the source is on GitHub.

Back to the blog