A browser that does things
Books meetings, fills forms, orders supplies — real work through real sites, driven by accessibility-tree automation instead of fragile CSS selectors.
An assistant that remembers,
acts, and stays yours.
A personal AI that runs on your machine — browsing the web, spinning up subagents, and building apps from chat. Every conversation makes it better at being yours. Your data never leaves.
QUICK ACTIONS
Browse the web
Search files
Run a research job
ACTIVE
Researching HybridDB v0.5.6
2 subagents running
Research the HybridDB v0.5 release
Subagent: research_hybriddb
Browser · scraped CHANGELOG.md and docs/API.md
App Builder · indexed 0.5.3 → 0.5.6 notes
Summarize what changed
Summary ready
0.5.6 fixes journal ordering and FTS backfill; 0.5.3 added
k_seeds to search_graph_ppr with graph-aware retrieval.
Our story
The assistant a doctor is allowed to have.
Most of the work doesn't need a frontier model — it needs a harness. This stack is built for the people whose data can't leave their machine, and it's being built in public.
The Open Assistants stack
One assistant, four building blocks.
Every layer is open source, MIT-licensed, and embedded — no servers to rent.
What it does Everything an assistant should do.
Everything an assistant should do.
Nothing it shouldn't.
Subagents, supervised
Parallel subagents on a SQLite work queue. Start them, course-correct them mid-run, cancel them when they're done.
Memory that compounds
99.9% session recall@5 on LongMemEval — zero LLM calls. It learns your style from every conversation.
Chat builds searchable apps
Describe a data app and get one. Every TEXT column gets keyword + vector search automatically.
Any model, your call
4,172+ models from 110+ providers through one SDK. OpenAI, Anthropic, Google, Ollama — no lock-in.
Billing that caps itself
Tenant pricing, tool-call budgets, and hard 402s the moment a budget is spent — usage-based billing built into the agent loop, with an owner dashboard for every run.
Local-first, always
Runs on your machine. No cloud, no telemetry, no training on your conversations.
Why it's different
An assistant, not another chatbot.
Measured on public benchmarks. Built for the machine that owns your data.
99.9%
Session recall@5
LongMemEval Oracle — 500 questions
zero LLM calls
98.4%
Session recall@5
LongMemEval S — 95.0% baseline + validated zero-LLM improvements (+0.034)
zero LLM calls
0.472 → 0.867
m@5 precision
After cross-encoder reranking — one lightweight local model
0%
Abstention rate
False-positive avoidance across every recall mode
zero LLM calls
Other assistants start fresh every conversation.
Yours remembers what you mean.
Other assistants only reply.
Yours does things — web, files, subagents.
Other assistants train on your data.
Yours keeps it on your machine.
Other assistants lock you into one model.
Yours runs any of 4,172+.
How it works
Talk. Act. Remember.
01
You talk
One request with full context — ask, delegate, or point at a problem.
02
It acts
97 tools across browser, files, shell, and web — parallel subagents, human approval for destructive steps.
03
It remembers
Every turn is embedded locally. Recall is hybrid keyword + vector, reranked — zero LLM calls on the happy path.
Free · Open source · Yours
Start with an assistant that gets you.
Runs locally, connects to the model you choose, and keeps your data on your machine.