RagLeap is a free platform for building AI employees — agents that learn from your business documents and talk to your customers on WhatsApp, Telegram, Discord, and voice. By default they take no action on their own. In approval mode, every real action waits for your yes first; full autonomy is opt-in.
# checks Docker, clones the repo, creates your .env $ curl -fsSL https://raw.githubusercontent.com/antonyrag/ragleap-core/main/install.sh | bash # asks which AI to use (Gemini, local Ollama or skip), generates your keys and starts the stack # manage it afterwards with: ragleap launch, stop, status, logs, update, key
$ git clone https://github.com/antonyrag/ragleap-core.git $ cd ragleap-core $ cp .env.example .env # add GEMINI_API_KEY (embeddings use Gemini by default); chat model via LLM_PROVIDER $ docker compose up --build -d
RagLeap
ragleap-ops, live-tested on a kind cluster. Kubernetes guide →RagLeap is open-source software you run on your own server. It is not a SaaS product, there is no RagLeap account, and no RagLeap-hosted database — your data is stored on infrastructure you control, and text is sent only to the AI providers you configure: your chat model and your embedding provider (Gemini by default).
Turns your documents — PDFs, policies, catalogs, FAQs — into AI employees that can answer questions accurately, hold a conversation across chat and voice channels, and optionally take real actions like sending an email or logging a task.
It doesn't act without permission. It doesn't browse the open web or run shell commands by default. It doesn't lock you into one AI provider, one vector database, or one deployment target.
Instead of a single assistant that tries to do everything, RagLeap ships with 46 pre-built roles — each with its own guardrails, memory, and skill tags — plus the tools to define your own. Route a request to the right one directly, or let a supervisor model pick for you.
/office, with an overview, approvals, employees by department, a task board, agent runs, an activity log and settings. It is reachable locally or through an SSH tunnel, by design.
A personal, always-on agent — broad reach across shell, browser and channels, built for one person's own machine. No concept of a business owner approving what it's allowed to do.
An orchestration layer that sits above agents like OpenClaw — org charts, budgets, governance. Closer to a management console than to a set of employees who actually do the work.
AI employees that already know your business — your documents, your policies — and a rule underneath every one of them: by default, anything that reaches outside a chat reply waits for you to approve it first. Full autonomy is an explicit opt-in, and sensitive roles can never run fully autonomously.
A chat reply is free. Anything that sends an email, posts to Slack, calls a webhook, or changes a setting is not — it stops here first, every time.
One action, from what you've configured — never a destination it invents itself.
A message goes to the exact channel and contact you set as the approver.
Checked against that one configured owner — a reply from anyone else changes nothing.
Only on YES. Every step is logged either way.
Legal, medical, tax and compliance-flavored employees are permanently held to semi-automatic or fully manual mode — including any role you create yourself.
A usage ledger records every LLM call an employee makes, with optional daily and monthly token caps — per employee or across the whole deployment.
Platform signatures are checked on every inbound message; an unsigned request is rejected rather than silently trusted.
Optional chain-of-thought and tree-of-thought modes, plus a self-correction pass that checks an answer against your documents before it's sent.
Send a request to the office as a whole — a supervisor model picks the right employee, or splits it across a small team and merges the results.
19 chat providers, including Gemini, Anthropic, OpenAI, Groq, Mistral and DeepSeek, local models through Ollama, and any OpenAI-compatible endpoint. You can set up a fallback chain for when a provider is down or rate-limited.
pgvector by default, or swap in FAISS, Pinecone, Weaviate, Qdrant or Milvus without changing how you call it.
WhatsApp, Telegram, Discord, and voice calls (Twilio, real-time transcription and speech), plus an n8n workflow integration.
Automated guard tests catch a half-added role before it ships unguarded, and a CONTRIBUTING guide walks through adding a new employee.
Every item below reflects the real state of the repo, not a projection. See the latest release notes for exact PR references.
Every contributor is pulled live from GitHub — including community fixes like improved short-query language detection. Open an issue before a large change; the project deliberately stays narrow in scope.
View all contributorsTwo public LinkedIn posts from people outside the maintainer team. We link to the originals so you can read them yourself.
AI Solution Builder and Full-Stack Developer
In a German-language post he names RagLeap Core among open-source alternatives, describing it as offering 46 role-based AI employees, self-hosted, MIT-licensed and with no license key. Full write-up: RagLeap Core: 46 built-in AI employee roles.
Highlighted by us. RagLeap does not vouch for the other claims in this post.
AI Intern at Gigin.ai
He describes his fifth merged open-source pull request as improving short-query language detection in RagLeap Core (PR #452). On his own benchmark of 32 short queries across 9 languages he reports langdetect at 43.75%, FastText at 90.63% and the final combined approach at 32 of 32.
Those are his figures on his benchmark, not a general accuracy claim for RagLeap. See his contributor page. Full write-up: how RagLeap detects the language of very short queries.
Read the full write-up: How RagLeap detects the language of very short queries. It covers Hardik's work in detail and mentions both posts.
Yes. It's MIT licensed and free to self-host. You pay only for the AI provider keys you use. Chat can run on a local Ollama model. Embeddings use Gemini by default (Google AI Studio issues free keys) and can be switched to Ollama, OpenAI, Mistral and other providers with EMBEDDING_PROVIDER.
By default, only with your approval. It can propose one action from tools you've configured, and in approval mode nothing runs until you reply YES from the exact channel and contact you set as the approver. Full autonomy is an explicit opt-in, and sensitive roles can never run fully autonomously.
No. There's no RagLeap account and no RagLeap-hosted database. It runs on your own server via Docker Compose.
Add one through the API at runtime, or contribute a built-in role — both paths are covered in the CONTRIBUTING guide, and a guard test stops a half-configured role from shipping without safety checks.
19 chat providers, including Gemini, Anthropic, OpenAI, Groq, Mistral and DeepSeek, local models through Ollama, and any OpenAI-compatible endpoint. Embeddings use Gemini by default and can be switched to Ollama, OpenAI, Mistral and other providers. You can configure a fallback chain between providers.
git clone https://github.com/antonyrag/ragleap-core.git
cd ragleap-core && cp .env.example .envPut your GEMINI_API_KEY in .env (embeddings use Gemini by default; set EMBEDDING_PROVIDER to use another provider). Pick the chat model with LLM_PROVIDER: 19 options including Anthropic, OpenAI, local Ollama, or any OpenAI-compatible endpoint. Then run docker compose up --build -d.
The app runs at localhost:8000. Ingest a file, then chat with any of the built-in employee roles.