Raven is an open-source team chat application built on the Frappe Framework — the same stack ERPNext runs on — that lives inside your ERPNext instance instead of requiring a separate Slack or Teams subscription. Since Raven v2, it also includes no-code AI agents that can create ERPNext records and read receipt images, using OpenAI's Assistants API behind the scenes. This guide on what is raven for erpnext is written for Indian SMEs, with code samples, ERPNext / Medusa recipes, and step-by-step fixes you can copy into a real project.
Raven started as a straightforward answer to a common gap: ERPNext teams needed internal chat,. Buying a separate Slack subscription meant paying for and maintaining yet another tool alongside the ERP itself. Raven closes that gap natively — channels, direct messages, document alerts tied to ERPNext records — all inside the same Frappe-based system.
What makes Raven worth a dedicated look now is v2's AI agent layer. Which turns it from "internal Slack alternative" into something that can actively act on your ERPNext data through conversation. I am Manoj, an ERPNext and Frappe implementation consultant at MithTech, and this is what Raven actually does today.
What does Raven actually do inside ERPNext?
Raven provides team chat — channels, direct messages and document-linked alerts — running natively inside an ERPNext instance. Since v2 it also ships no-code AI agents. Those agents read instructions in a chat message and act on ERPNext data in response. The chat layer alone replaces a standalone Slack or Teams subscription for internal communication; the AI agent layer is the part that's newer and less widely understood.
Document alerts are worth calling out on their own. This is not a generic notification system. Raven posts to a channel when a specific ERPNext document changes — a Sales Order submitted, a Purchase Invoice approved. The conversation stays tied to what is actually happening in the ERP, rather than a separate activity feed nobody checks.
Skimmable summary: Raven is native team chat for ERPNext — channels, DMs and document-linked alerts. Since v2 it also has no-code AI agents that act on ERPNext data from a conversation. Two genuinely separate capabilities in one app.
How do Raven's AI agents work?
Raven's AI agents connect to OpenAI's Assistants API behind a no-code configuration layer. You define an agent with specific instructions and a defined task. In response to a chat message, that agent can create or update ERPNext records. It can also read a receipt or invoice image and draft the matching document. No custom code is required to set up an agent; the configuration happens through Raven's interface.
Permissions are enforced at the user level, not the agent level: an agent only has access to what the person operating it in that conversation is already permitted to see or create. If a user can't create a specific document type, the agent acting on their behalf can't either — this scoping is what keeps agent permissions from silently exceeding a human user's actual access.
Skimmable summary: Raven AI agents run on OpenAI's Assistants API through a no-code setup. They can create records and read receipt or invoice images. They inherit the permissions of whichever user operates them — there is no separate agent-level access.
Can Raven run without sending data to OpenAI?
Raven can run without sending data to OpenAI or Anthropic. It uses a local LLM through Ollama, self-hosted on your own infrastructure. That includes a server you control yourself, such as an AWS instance you manage. This is the option that matters for teams where data privacy is a hard requirement. Since the AI agent then runs entirely against a model you host rather than a third-party API.
The trade-off is real, and worth stating plainly. Self-hosted open models available through Ollama are generally less capable than the latest cloud-hosted Claude or GPT releases. Choosing local-only buys privacy with capability. It is not a free substitute. Testing against your actual use cases before committing either way is worth the time.
Skimmable summary: Raven's Ollama support runs AI agents against a self-hosted local model instead of a cloud LLM. You trade some capability to keep all data on infrastructure you control. That is a real trade-off, not a strictly better option.
How do I set up Raven on my ERPNext instance?
You install Raven as a Frappe app on your existing ERPNext bench, or through the Apps tab on Frappe Cloud. Channels and direct messages work immediately, with no further configuration. Enabling the AI agent layer is a separate step, and worth doing in this order:
- Install the app.
bench get-app raven && bench --site <site> install-app ravenon a self-hosted bench, or the Apps tab on Frappe Cloud. - Confirm chat works first. Create a channel, send a message. If this does not work, no amount of LLM configuration will help.
- Wire the document alerts you actually want. Start with one or two — a submitted Sales Order, an approved Purchase Invoice. Alerting on everything trains people to ignore the channel.
- Connect a model. An OpenAI API key, or Ollama pointed at a local model.
- Define one agent, scoped narrowly. Give it instructions and restrict it to the DocTypes it needs. Read-only to begin with.
- Test it on staging. Have it draft documents you check by hand before anything is granted write access.
- Grant write access last, one DocType at a time, once the drafts have been right for a while.
On step 4, the choice of model is a real trade-off rather than a preference:
| Where your data goes | To OpenAI's API | Stays on infrastructure you control |
| Capability | The latest GPT releases | Generally behind the newest cloud models |
| Running cost | Per-token API billing | Your own compute, no per-token fee |
| Hardware needed | None | A machine that can hold the model |
| Setup effort | An API key | Install, host and maintain the runtime |
| Best when | Capability matters more than residency | Data residency or policy rules out a cloud LLM |
Start read-only
As with any AI agent given write access to live ERPNext data, start an agent's permissions narrow and read-only, confirm its behaviour against real test cases on a staging site, then widen access deliberately. This applies to Raven's agents the same way it applies to any MCP-based AI connection into ERPNext.
Skimmable summary: Raven installs as a standard Frappe app, and chat works immediately. AI agents need a separate LLM connection and explicit DocType scoping. Test them on staging before granting write access on a live instance.
Want Raven set up inside your ERPNext instance?
I set up Raven's chat and AI agent layers as part of ERPNext implementations — scoped permissions, staging tests, and a model choice (cloud or self-hosted) that fits your data-privacy requirements. See my broader ERPNext AI features guide for how Raven fits alongside other AI options.
Frequently asked questions
+Is Raven free to use with ERPNext?
Yes — Raven is free and open source, with no licence fee, consistent with the rest of the Frappe ecosystem. If you connect a cloud LLM like OpenAI for the AI agent features, you pay that provider for usage; Raven itself doesn't add a cost.
+Do I need coding experience to set up Raven's AI agents?
No — agent configuration in Raven is designed to be no-code, done through Raven's own interface rather than requiring custom Frappe app development. Connecting an LLM API key and scoping agent permissions to specific DocTypes is a configuration task, not a development one.
+Can Raven agents access data I don't have permission to see?
No — Raven agent permissions are scoped to the user operating the agent in a given conversation. If that user doesn't have access to create or read a specific document, the agent acting on their behalf doesn't either, which limits how much an agent can do regardless of how it's instructed.
+Is Raven the same as ERPNext's other AI features, like Frappe Assistant Core?
No — they're separate tools solving different problems. Raven is chat-native, with AI agents that act inside a conversation; Frappe Assistant Core connects ERPNext to Claude or ChatGPT over the Model Context Protocol for natural-language querying and reporting outside of a chat interface. Many ERPNext AI setups use both together rather than choosing one.
About the author
Manoj is an ERPNext and Frappe implementation consultant at MithTech in Bengaluru, and sets up Raven's chat and AI agent capabilities as part of ERPNext implementations for clients evaluating practical, scoped AI use cases.