KIAgent

[KIAgent logo]

KIAgent

Your personal AI brain — local, open source, speaks MCP.

Website · Download · Architecture docs

[License: MIT] [Platforms]

[KIAgent main window — Sources screen with Gmail and HubSpot syncing]

Everyone is building the company a brain — a knowledge layer that ingests everything an organization knows, so AI can answer and act with real context. KIAgent is that layer for you personally. A company brain lives in someone else's cloud and belongs to your employer; your personal AI brain runs on your own machine, works with whatever AI you choose, and is yours forever.

Concretely, KIAgent is a desktop app that ingests your personal data — mail, documents, chats, notes — on your own machine, stores it in a local database, and serves it to AI assistants over MCP. Everything stays on your computer: ingestion, parsing, OCR and vision all run locally — with on-device inference powered by llama.cpp — so your AI assistant can know your world without your data ever leaving your machine.

At its core, KIAgent is a platform: a small set of host capabilities (MCP, database, local LLM, filesystem, web access) wired together by an ingestion engine, plus a plugin/extension system that lets sandboxed connectors use those capabilities to bring in new data sources.

flowchart TB    GMAIL(["Gmail"])    DRIVE(["Drive"])    SLACK(["Slack"])    NOTION(["Notion"])    MORE(["…"])
    PLUGINS["Plugins / extensions"]
    WEB["Web"]    LLM["Local LLM"]    FS["Filesystem"]    DB[("DB")]    MCP["MCP"]
    GMAIL --> PLUGINS    DRIVE --> PLUGINS    SLACK --> PLUGINS    NOTION --> PLUGINS    MORE --> PLUGINS    PLUGINS <--> WEB    PLUGINS <--> LLM    PLUGINS <--> FS    PLUGINS <--> DB    DB --> MCP
    classDef service fill:#e8f0fe,stroke:#4285f4,color:#1a3c78    classDef core fill:#e6f4ea,stroke:#34a853,color:#1e4620    classDef hub fill:#fce8e6,stroke:#ea4335,color:#7a1c14,stroke-width:2px
    class GMAIL,DRIVE,SLACK,NOTION,MORE service    class WEB,LLM,FS,DB,MCP core    class PLUGINS hub

Plugins sit at the center: each one connects to a third-party service (Gmail, Drive, Slack, Notion, ...) or the local filesystem, pulls your data in, uses the local LLM to process it, and lands it in the database — which MCP then serves to AI assistants. Access is mediated and permission-gated; plugins never touch these capabilities directly.

This repo: kiagent-core

This repository is the MIT-licensed core that KIAgent is built from, in the same spirit as VS Code and Code-OSS:

• kiagent-core (this repo) — the open-source core. npm run package:oss produces an unbranded kiagent-core build you can run and redistribute under the MIT license.
• KIAgent — the branded, signed product built from this core and distributed at localkiagent.com/download.

OAuth-backed sources (Gmail, Microsoft) read client credentials from the environment at build time — see .env.example / CI secrets. Forks and OSS builds supply their own OAuth client IDs.

Main components

• Ingestion engine (src/main/core/engine) — pulls batches from sources on a cadence, converts raw items (mail, PDFs, images) into indexed documents, and commits them to the store. Built-in sources live in src/main/sources (Gmail, IMAP, Microsoft 365, local folders).
• Database (src/main/core/store, src/main/db) — a SQLite corpus (better-sqlite3) holding documents, metadata and search indexes, written by the app and read by the MCP processes.
• MCP server (src/main/core/mcp, src/main/mcp) — exposes the corpus to AI assistants over two transports sharing one tool registry: an HTTP server inside the app, and a standalone stdio entry point that clients like Claude Desktop spawn directly.
• Local LLM (src/main/providers/local-llm, src/main/providers/apple-vision) — on-device inference via a bundled llama.cpp server (backend auto-detection, curated model tiers) plus native OCR/vision helpers for scanned documents and images.
• Filesystem — local-folder ingestion, model storage, temp workspaces for conversion workers (src/main/workers, src/main/converter).
• Web — outbound HTTP for source APIs, OAuth flows (src/main/auth, src/main/platform/oauth-providers.ts), and marketplace downloads.
• Plugin / extension system (src/main/platform, src/main/marketplace) — connectors run in isolated host processes with a manifest-declared, permission-gated view of the platform (sources, auth, storage, network). A GitHub-backed marketplace handles discovery, install and updates. A second, privileged tier lets a product build ship first-party extensions inside the app package itself — see docs/architecture/extension-platform.md for the full model, including the bundled/unsafe.mainProcess tier and product.json.
• Renderer (src/renderer) — the React UI: source setup, marketplace, MCP connection status, settings and logs.

Developing

npm installnpm start          # run in development (hot reload)

Other useful scripts:

npm test                 # jest test suitesnpm run lint             # eslint + prettiernpm run typecheck        # tsc, no emitnpm run package          # build a distributablenpm run package:oss      # build an unbranded kiagent-core distributablenpm run vendor:inference # fetch llama.cpp server + build vision helper

Repository layout

src/main       Electron main process: engine, store, MCP, platform, sources, providerssrc/renderer   React UI (screens: Sources, Marketplace, Connection, Settings)src/shared     Contracts and UI primitives shared across processesconcept/       Types-only blueprint for the core redesigndocs/          Design specs and implementation plansnative/        Native helper sources (vision/OCR)scripts/       Build and vendoring scripts

MCP directories

KIAgent is published on the official MCP registry as com.localkiagent/kiagent (see server.json) and listed on Glama:

[kiagent-core MCP server]

Acknowledgements

KIAgent's on-device inference is powered by llama.cpp, created by Georgi Gerganov and maintained by the ggml-org community (MIT license). The app bundles prebuilt llama-server binaries from the official llama.cpp releases (npm run vendor:inference fetches them at build time). Running capable models privately on consumer hardware is only practical because of their work — thank you.

License

MIT — see LICENSE.