AI agents that
live on your machine.
Pumkin is a local-first agent IDE for developers. Build, run, and audit AI agents using your own Ollama models and your own MCP tools. No cloud. No API keys. No recurring fees.
Real output. Local Ollama. llama3.2:3b. Built-in get_time tool. 1.5s.
What's in the box.
Everything below is built, tested, and shipping in v0.0.1. No roadmap items in this list — what you pay for is what you get.
Native desktop app
Tauri-based installer for Windows (macOS and Linux next). Bundled Node.js — no system dependencies beyond Ollama itself.
Local models via Ollama
Talks to your existing Ollama install. Picks up your installed models automatically. Run llama3.2:3b on 8GB of RAM or qwen3:14b if your machine can handle it.
MCP tools, allowlisted per agent
Connect any stdio or Streamable HTTP MCP server. Pumkin runs each tool through a JSON schema check and per-tool timeout. Tools are namespaced so the model can't accidentally call something it shouldn't.
Per-call approval gate
Mark any tool as 'requires approval' and Pumkin pauses the run, shows you the exact call and args, waits for your green light. Approve once. Deny once. No global trust.
Run history with full event replay
Every run persisted to SQLite. Inspect the model response, the tool calls, the tool results, the timings. Replay anything. Debug what the model actually saw, not what you thought it saw.
Multi-turn conversations
Persistent threads with full context. Ask a follow-up; Pumkin remembers turn one. Same agent, multiple conversations, all logged.
Pumkin is for you if…
- You already run Ollama and you've been wiring agents together by hand
- You don't want your prompts, tool calls, or conversation history leaving your machine
- You're tired of paying $20/mo to five different AI products to do what one local model could handle
- You want to inspect every tool call your agent makes before it runs
- You've used Open Interpreter, AnythingLLM, or LM Studio and wanted them to do more, not less
- You believe a tool you own outright beats a subscription you rent forever
Pumkin is not for you if…
- You don't have Ollama installed and don't want to install it
- You want to chat with one model — Pumkin is for agents that do things
- You need the absolute best model quality — local models can't match GPT-5 on hard reasoning
- You're looking for a no-code product. Pumkin is built for people comfortable in a terminal.
- You expect Mac/Linux support today. Windows first; the rest is on the roadmap.
Honest answers.
How is this different from LM Studio or AnythingLLM?
Different problem. LM Studio is a model runner — it lets you chat with a local model. AnythingLLM is a RAG product — it lets you ask questions over your documents. Pumkin is an agent IDE — it lets you build named agents with specific tools and audit every call they make. If you've tried wiring agents together by hand with the Ollama API and MCP servers, you know how much glue this saves.
What do I need installed on my machine?
Just Ollama (free, from ollama.com). Pumkin's installer bundles Node.js and everything else. No Rust, no Visual Studio, no Python. Run the installer, you're done.
Is Pumkin open source?
Yes — Pumkin is free and open source under the MIT license. The full source is on GitHub: read it, modify it, fork it, even ship your own version. No license key, no account, no paid tier. It runs entirely on your machine.
Mac and Linux support?
Windows installer ships today. Mac and Linux are next — the Tauri shell is cross-platform; the bottleneck is the bundled-Node and code-signing story on each platform. They'll be free when they land, same as everything else.
What if it doesn't work for me?
Open an issue on GitHub or email me. I'm one person, but I read everything and I'll try to help. And because it's MIT-licensed, you can always dig into the source and fix it yourself — PRs welcome.
What about updates?
Updates are free, always. New releases land on GitHub — watch or star the repo to get notified. No paid upgrades, no version gating.
Why is it free?
Because I'd rather have people actually using it than guard a small paywall. Pumkin is part of how Software For Humans builds in the open — free, MIT-licensed, and genuinely useful. If it saves you a weekend of wiring agents together, that's the point.
Who built this?
An indie founder under Software For Humans, LLC, based in Virginia, USA. A track record of building, shipping, and supporting small tools. Not a venture-backed startup. Not a side project of a bigger company. Just one person who wanted this thing to exist.
Get Pumkin.
- Windows installer today; macOS and Linux next, all free
- Full source on GitHub — read it, fork it, ship your own
- No account, no license key, no telemetry
- Bundles Node.js; just install Ollama and go