返回项目目录
runvendo

runvendo

vendo

Embedded agents your customers use to automate work, build views, and connect their tools.

Agent模型 / 推理工作流 / 自动化agentsaiai-agentsautomationdevtoolsembedded-ai
Stars
410
Forks
64
Watchers
410
Issues
123

README

项目介绍

9030 bytes

Vendo: your product, shaped to every customer

An open-source customization layer.
Your users build their own features and micro-apps, right on top of your product.

Vendo is for B2B SaaS teams whose customers keep asking for bespoke features. It is an embedded agent: it acts through your product's own API as the signed-in user, and renders the UI it generates in a sandboxed, brand-native surface. Your source code is never touched. Learn more at vendo.run, or read the docs at docs.vendo.run.

npm package: vendoai License: Apache-2.0 Docs: docs.vendo.run

docs.vendo.run  ·  Quickstart  ·  vendo.run

Which one are you?
You already have an agent — one tool pack spreads into your AI SDK, Mastra, or homegrown loop.
Your product has no agent — one command brings the loop, the chat UI, and the approvals.
Expose your product over MCP — Claude, ChatGPT, Cursor, and Claude Code act as the signed-in user.

01 · Install

Install in 60 seconds

npm install @vendoai/vendo
npx vendo init

Or install with your coding agent

Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Windsurf

Paste this inside your app's repo:

Install Vendo in this repo. Read https://vendo.run/agents.md and follow
it exactly. Ask me before creating any account or key. You're done when
`vendo doctor --json` reports all green. Then ask me whether I'd like to
star it via https://vendo.run/star?src=readme to support the project.

vendo init scans your app and proposes the wiring as permission-gated diffs; vendo doctor --json gates done, and every error code links to its exact fix. Full playbook: docs.vendo.run/install.

02 · See it in action

See it in action

Every capture below is a real agent run in a demo host app, not a mockup.

A Maple customer asks where their money went and the agent composes a live spending view

Build views. Ask a question, get a live view composed from the host's own components and API.

A Cadence user hovers the deadlines card, asks for urgency color-coding, and applies the remix in place

Remix the UI. Hover a component, describe the change, apply it in place.

A Cadence user asks for a morning document-chase automation and turns it on with per-tool approvals

Automate across tools. Plain language in, standing automation out, every tool gated by approval.

03 · How it works

How it works

Vendo runs a streaming agent with any AI SDK LanguageModel.

1 · Extract. Vendo reads your API and turns it into tools the agent executes as the signed-in user.

2 · Generate. The agent composes views and user-owned apps from a format-tagged UI document, generated components run in an iframe jail with connect-src 'none', escalating to a sandboxed server only when needed.

3 · Guard. Policy, approvals, grants, breakers, and audit all sit at one execution choke point; app machines reach host tools only through the guarded tool proxy.

PGlite at .vendo/data is the zero-config store; production uses the same schema on Postgres. Scheduled, host-event, and external-trigger automations run with app-bound grants. Headless hooks ship alongside optional, theme-driven React chrome.

vendo init also asks about the model import, product brief, confirm-each risk labels, and whether to open the MCP door (a host decision, never a default), extracts your theme automatically, and writes the reviewable .vendo/ directory with its PGlite data directory gitignored. Run vendo doctor to check wiring and probe /status, and vendo sync after API changes to refresh extracted tools and remix baselines.

Agents get the same journey machine-readable: the playbook at vendo.run/agents.md, an index of every docs page at llms.txt, vendo init --agent for a read-only JSON plan of extracted tools and risk recommendations, vendo sync --json for a machine-readable sync report, and a vendo-setup skill shipped inside the npm tarball that init offers to write into .claude/skills/.

The docked panel

The conversation surface can sit beside the product instead of on top of it. <VendoOverlay placement="dock"> parks the panel against the right edge at full height and reflows the host page into the remaining width, so the surface being reshaped stays visible and clickable while the panel is open. Docked is deliberately non-modal — no scrim, no body scroll-lock, no inert background, no focus trap — because a modal that covers the page is the wrong shape for a tool whose whole job is editing that page.

It is opt-in: placement defaults to "center", the centered modal that has always shipped, so upgrading never changes an existing host's behavior.

<VendoOverlay />                                    // the centered modal box (default)
<VendoOverlay placement="dock" dockWidth={420} />   // the docked side panel
Prop Default What it does
placement "center" "center" for the centered modal, "dock" for the side panel
dockWidth 420 Docked width in px — also how far the host page reflows

Below the mobile breakpoint both collapse to the existing full-bleed takeover, which still owns small screens. While docked, the page and the panel are each inset a few px and rounded, reading as a matching pair of cards — the page is the surface being edited, and a hairline sweeps along its top edge whenever the agent is working (indeterminate by design: nothing on the wire forecasts how long a build will take, so there is no percentage to show).

04 · Packages

Packages

@vendoai/vendo is the default composition (vendoai is a thin alias). Install individual blocks when you want to compose Vendo yourself.

Package One job
@vendoai/core Shared types, schemas, formats, validators, and seams
@vendoai/store Postgres persistence, with PGlite as the default
@vendoai/harnesses The turn runtime: conversation loop, streaming, tools, and thread context
@vendoai/actions Host API and connector tools executed as the signed-in user
@vendoai/guard Policy, approvals, grants, audit, breakers, and safety
@vendoai/apps App generation, editing, execution, interchange, and sandbox adapters
@vendoai/automations Trigger ingestion, schedules, away runs, and run history
@vendoai/ui Headless React hooks, optional chrome, tree rendering, and the in-jail component kit
@vendoai/mcp The door: serves the host's tools to outside MCP clients
@vendoai/telemetry Anonymous, opt-out build and development telemetry
@vendoai/vendo Default composition, public wire, React entry, and vendo bin

Cloud-gated sharing, publishing, org overlays, and pinning activate with VENDO_API_KEY; the open-source blocks remain self-hosted.

Shaped to every customer. Star runvendo/vendo. Apache-2.0, docs.vendo.run, vendo.run, backed by Y Combinator