bench get-app https://github.com/MalikZu/frappe_agents

[![Listed on Frappe Gems](https://frappegems.com/api/method/frappe_gems.seo.badge?app=MalikZu%2Ffrappe_agents)](https://frappegems.com/gems/apps/MalikZu/frappe_agents)

An agent is a record you create in your site, not a service you deploy. The permission model is the product. ## Why this one Most ERP chat assistants are a text box bolted onto a service account with broad rights. This app is built the other way around: - **An agent has no rights of its own.** Every run is bound to a real user and can never see or do more than that user could by hand. - **Agents draft. Humans submit.** No code path submits, cancels, or deletes without a proposal that a *different* person approves — applied under the approver's own permissions. - **Every call is on the record.** Each tool call writes an audit row — on success, denial, and error — that nobody can edit or delete. A shipped report shows whether reviewers actually review or just click approve. - **Document text is data, never instructions.** Everything people wrote — comments, emails, field values — reaches the model wrapped as untrusted. - **One switch turns it all off.** A kill switch a running job actually sees. ## What you get - **Agent Chat** and an **Ask Agent** panel on any document — with conversation history, per-conversation model choice, and live tool progress. - **Approval queue** for anything that commits the business, with separation of duties and an edited-before-approval quality metric. - **Document extraction**: an attached PDF or image becomes a draft for human review, with bank and payment details held back until confirmed by hand. - **Any model provider** over two wire formats (OpenAI-compatible and Anthropic) — OpenRouter, Ollama and friends included. The provider is a config field, not a dependency. - **Zero pip dependencies.** The agent loop is a vendored, tested harness; nothing new enters your bench's shared environment. ## Requirements - Frappe v16 - Python 3.14+ ## Install Early software — expect sharp edges and breaking changes. On a real site, pin a release: `main` moves whenever something lands. ```bash bench get-app --branch v0.5.0 https://github.com/MalikZu/frappe_agents bench --site yoursite install-app frappe_agents ``` Installing creates the four roles (Agent Manager, Agent User, Agent Approver, Agent Auditor) and registers the built-in tools. Then, in the Desk: create an LLM Provider, a Model Profile, and an Agent — and open Agent Chat. A provider's **Base URL** is the host that receives your prompts and your API key, so it must be an `https://` address on the public internet. To point at a model you run yourself — Ollama on the same box, something on your own network — tick **Self Hosted** on the provider. That checkbox is what allows `http://` and addresses like `localhost` or `10.x`, and it is checked again before every request. Redirects are never followed. ## Docs - [Agent Chat](docs/chat.md) — conversations, the composer chips, and the document a conversation is about. - [Files in chat](docs/files-in-chat.md) — naming a file an agent may read, uploading one, and where uploads live. - [The agent harness](docs/harness.md) — the vendored run loop. ## License MIT