Pumkin

A local-first IDE for building and running AI agents against your own Ollama models. No cloud inference, no API keys, no per-token cost — everything runs on your machine.

Free and open source under the MIT license. Grab a build from Releases, or build from source below.

Heads up: the npm package and the SQLite database are still named agentkit internally — that rename is in progress and doesn't change how anything runs.

Quickstart (web UI)

# Root: agentkit runtime
npm install

# Web app
cd web
npm install
npm run dev          # http://localhost:3000

Then in the UI: create an MCP server (optional) → create an agent → click into it → enter a prompt → see the event stream live. The SQLite DB lives at ../agentkit.db relative to web/ (override with AGENTKIT_DB).

Quickstart (CLI)

ollama pull qwen3:8b
npm run cli -- agents:create          # NAME=... MODEL=... SYSTEM_PROMPT=... BUILTIN_TOOLS=get_time
npm run cli -- run <agentId> "what time is it?"

Requires Ollama running locally (default http://localhost:11434) with a tool-capable model pulled. Requires Node 22.5+ for the built-in node:sqlite.

Layout

agent.ts           runtime: agent loop, Ollama tool-calling, safety gates (zero deps)
mcp.ts             adapter: MCP servers → agent tools (only file that pulls the MCP SDK)
store.ts           SQLite persistence: agents + mcp_servers tables (node:sqlite, no native deps)
loader.ts          turns a stored agent row into a live Agent with tools wired up
builtins.ts        registry of stock tools (get_time, read_text_file, ...)
cli.ts             terminal CRUD + run command
example.ts         minimal runnable demo, no DB
store-test.ts      end-to-end integration test against a real MCP server
web/               Next.js 15 UI over the same store + runtime

Architecture

Four layers, built so the data model flows downhill:

1. Runtime (agent.ts) — agent loop with allowlist, schema validation, per-tool timeouts, approval gate, event stream.
2. MCP adapter (mcp.ts) — brings the whole MCP ecosystem in. Stdio for local servers, Streamable HTTP for remote. Tools namespaced (server__tool) and gated by default.
3. Store + CLI (store.ts + loader.ts + cli.ts) — SQLite persistence + a runtime-config loader + a usable terminal product.
4. Web UI (web/) — Next.js 15 App Router. Server actions for CRUD, an SSE endpoint streaming agent events to the browser live. Distribution later via Tauri wrapping the same code.

Models

Local agents live or die on tool-calling reliability. Skip anything under 8B for real use, and prefer models specifically tuned for it.

• qwen3:8b — current dev pick, reliable tool calling, ~5GB.
• llama3.1:8b — also reliable, similar size.
• qwen3-coder:30b-a3b-instruct-q4_K_M — MoE, ~18GB, faster inference than dense 30B.
• qwen3-coder:32b — strongest at this scale (24–32GB VRAM at q4/q5).
• llama3.3:70b — safest if you have 48GB+ VRAM.

Models smaller than 8B (or earlier 7B variants like qwen2.5-coder:7b) often imitate tool-call JSON in the response text instead of using Ollama's structured tool_calls channel. The runtime correctly ignores those — a security feature, not a bug — but the agent will appear to "not call tools." Pull a better model.

Status

Runtime, MCP adapter, SQLite store, loader, CLI, and web UI complete and integration-tested. Next: distribution via Tauri; agent edit page; live model picker via Ollama's /api/tags; interactive approval UI.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Report security issues privately via SECURITY.md, and please follow the Code of Conduct.

License

MIT © 2026 Software For Humans, LLC.