v0.1.0 alpha · BSL 1.1 · self-hosted

The self-improving personal AI you own.

An open-source personal AI that runs on your machine, reaches you on Signal, and validates its own nightly improvements against tests it cannot cheat. Single-user by design. Sandboxed plugins. 12,600 lines you can read in an afternoon.

Self-host on your Linux server. Read the install script first.

Join the Kwisai Cloud waitlist

Hosted, managed, $49/mo + tokens. Coming soon.

[ 30-second demo: Signal voice note → Kwisai response ]

Demo video lands when launch artifact is recorded.

Live dogfooding metrics

…

GitHub stars

live

…

Self-hosters

opt-in telemetry

…

% Kwisai written by Kwisai

self-improvement loop

…

Nights survived on maintainer's box

continuous uptime

All numbers will go live from the /open dashboard.

What is different

Three things no other personal AI does today.

Learns from your conversations, not someone else's.

Every night, Kwisai proposes code changes to its own behavior, replays them against a golden dataset of your past conversations, and accepts or reverts on a six-dimension empirical score. Every change lands in git. Your AI gets better because you used it, not because a team on the other side of the world retrained a shared model.

/why §5

A bad plugin cannot reach your digital life.

Every third-party plugin runs in an OS-level sandbox: a child process with a blocked filesystem, a blocked network, and a capped memory budget. SES compartments, bwrap, and cgroups, stacked. The worst case is a locked room, not arbitrary code on your box.

/why §4

Three agents, and a router that learns.

Kwisai is three specialist agents: a fast Haiku for simple things, a Sonnet for code work, an Opus for deep research. A three-tier router decides which agent gets each message, and a daily cron teaches the cheap tiers to handle more of tomorrow's traffic. Ninety percent of what you ask costs pennies.

/why §2

Thesis

Personal AI should run on your machine, be reachable through the messaging app you already use, improve itself nightly against tests it cannot cheat, and stay out of the cloud, including ours.

Read the manifesto

Two paths

Run it yourself, or wait for us to run it.

No guilt about either choice. The same Kwisai, two operating models.

Comparison of self-hosted and Kwisai Cloud options
Attribute Self-host Kwisai Cloud
Price $0 software · ~$5–20/mo VPS$49/mo + tokens at provider rates
Hardware You bring a Linux serverIncluded
Setup time ~2 hours~2 minutes
Where data lives On your machine, in SQLiteDedicated EU instance we operate
LLM tokens Your API keyYour API key, billed at cost
Updates git pullManaged
License BSL 1.1, source visibleCommercial use granted

Read every line

12,600 lines of TypeScript. About 10 npm dependencies. No frameworks.

Auditable in an afternoon. Forkable in a weekend. The plumbing is ordinary Node.js: an Express-flavored gateway, BullMQ queue, SQLite, Pino logging. The interesting part is the architecture.

kwisai/
├─ gateway/         Node.js + TypeScript core
│  ├─ src/agents/   3-tier router · adapters
│  ├─ src/plugins/  SES compartments · RPC
│  ├─ src/jobs/     Self-improvement loop
│  ├─ src/queue/    BullMQ + Redis
│  └─ src/db/       SQLite session store
├─ webui/           SvelteKit chat client
├─ caddy/           Reverse proxy + TLS
└─ scripts/         install.sh · backup · audit

Honest answers

Frequently asked.

Is this ready for production?

No. v0.1.0 alpha. Run it on a server you can reformat without crying. The Cloud waitlist is for people who want to wait until it is.

Why BSL 1.1 and not MIT?

Each version converts to Apache 2.0 after two years. BSL prevents third parties from offering Kwisai Cloud as a hosted service while we build the business. You can self-host commercially, fork it, study it, contribute to it. Read the license page.

What does "self-improving" actually mean?

A scheduled job analyzes recent conversations, proposes code changes, runs them against a golden dataset of past interactions, scores them on cost, conversation quality, error rate, and four other dimensions, and either commits or reverts. You see every change in git. You can pause or veto the loop at any time.

What runs on my server, what runs in the cloud?

On your server: the gateway, plugin sandboxes, message queue, session store, voice transcription (Whisper), the self-improvement loop. In the cloud: only the LLM provider you choose (Anthropic, OpenRouter, or your own Ollama instance). Your conversations never leave your machine except as prompts to the LLM you configured.

Who is this for?

Sysadmin-developers. People comfortable with systemd, Docker is optional, Caddy, journalctl. If you do not run your own server today, Kwisai Cloud will be friendlier when it ships.

Linux only?

Linux only. Tested on Ubuntu 24.04 LTS. Other distributions should work; macOS and Windows are not supported and are not on the roadmap.