sunday

Sunday

S.U.N.D.A.Y. — Sees · Understands · Notices · Decides · Acts · Yours

Can any model go AGI, given the right harness?

Sunday is an open experiment to find out. It is a local-first agent harness: a voice and screen front end, a memory, a skill library and an action layer wrapped around a swappable model. The bet is that a lot of what looks like "intelligence" comes from the loop around the model (planning, verifying, remembering, learning from failure) and not only from the weights. If that is true, a small local model with a good harness should keep getting more capable, and a bigger model should get further still.

We don't claim to have AGI. We claim a measurable target, and the repo is organised to chase it:

v1 bar: Sunday completes open-ended, multi-day goals on your own machine with no hand-holding, scoring >= 70% on a fixed 50-task eval suite, with the pass rate rising week over week while the model stays constant.

If the number moves, the harness is doing real work. If it doesn't, we learn that too. Either way it is a number, not a vibe.

Track progress in the v1 master issue. Docs: sunday.rithul.dev.

How it fits together

flowchart LR
    in["Voice, typing,<br/>screen"] --> route["Router<br/>one typed Action"]
    route --> talk["Teach, chat,<br/>research, memory"]
    route --> agent["Agent loop<br/>plan, act, verify"]
    agent --> os["Your computer<br/>behind approvals"]
    talk --> model["Swappable model<br/>Ollama, MLX, hosted"]
    agent --> model
    os --> log[("Audit log")]

Every box is a module with a test; the architecture page has the loop, the approval tiers and the data model in detail.

Where it is today

Honest status. This is early software, macOS only, and runs on a 4B local model by default.

Sees	Screenshot, local OCR with exact boxes, vision model for diagrams, click-through overlay that points at things
Understands	Teaches any subject (plans, concepts, quizzes, spaced review), imports your notes
Notices	Wake word, voice activation, your voiceprint, long-term memory with provenance
Decides	Every utterance becomes one typed Action; a deterministic rules router, plus an optional on-device decision model (fine-tuned Laya) and a memory of your corrections for what rules can't match
Acts	Opens apps, pages and files, types, presses shortcuts, clicks. Checks each step really happened, retries what didn't, and stops honestly. All behind tiered approvals and an audit log
Yours	Ollama, MLX on this Mac, or a hosted model for chat or planning; SQLite on disk; web access off unless you switch it on
Measured	An eval suite with a sandbox, routing and agent tasks, and a prompt-injection red team, run on every PR

What each of these does in detail: docs/FEATURES.md.

What's next (v1)

1. Measure: grow the eval suite to 50 tasks and track the pass rate every night.
2. Agent loop: persistent tasks that survive restarts, and follow-ups on what is already open (the loop and verifiers have shipped).
3. Tools: MCP, a sandboxed shell, a browser driver.
4. Memory that compounds: successful runs become reusable skills, gated by the evals.
5. Self-improvement: failure analysis and a promotion gate; the router already learns from corrections.
6. Everywhere: Windows and Linux.

The full list, with status, is the roadmap. Everything is a labelled issue: browse roadmap.

Quick start

scripts/fetch-models.sh     # speech models into ~/.cache/sunday/models
scripts/build-ocr.sh        # screen helper; needs Xcode command line tools
ollama pull qwen3:4b-instruct && ollama pull nomic-embed-text-v2-moe
ollama pull qwen3-vl:4b-instruct   # only for diagrams and non-text screens
bun install && bun run tauri dev

Check it with cd src-tauri && cargo test (about a second, needs nothing installed) and bun run test:live for the tests that need Ollama and the models.

Say "sunday" or hold fn, then talk, or type in the workspace. Full walkthrough and troubleshooting: docs/GETTING_STARTED.md.

Docs

• Getting started: install, first run, things to try, where to begin hacking
• Architecture: the agent loop, approvals, routing, platform layer and data model, as diagrams
• Roadmap: what shipped and what is next
• Features: what exists and how it behaves
• Providers: Ollama, MLX and hosted models, keys and privacy
• Router: how an utterance becomes an action, and how that gets smarter
• What's in the repo: every module, the data flow, the database tables
• Testing and Evals: hermetic tests, live tests, the eval suite and CI
• Contributing: the issue-to-PR flow, safety rules, working with AI coding agents

Help build it

Pointing your own Claude (or any coding agent) at this repo is welcome. Start from issues labelled claude-friendly, good first issue or help wanted. Windows and Linux ports need people on those machines.

License

GPL-3.0-or-later. Forks and derivatives must stay open source under the same license. See LICENSE.