aiti-agent — the Main Web Battleground

In one sentence: a self-hosted personal AI development workbench powered by pi — the browser as the primary operating surface (same-source compatible with pi: pi CLI / pi IDE extensions operate the same set of sessions), one-command docker run startup, just get to install tools. Backstory: built on top of the open-source project pi-web; the deeper it was used and the more it was modified, the more it grew into its own release line (see Project Origin).

1. Positioning

Dimension	Conclusion
What to build	Our own AI development workbench (not "yet another pi-web", but "use pi to build our own tools")
Foundation	Built on the open-source project agegr/pi-web (incl. the ecosystem's only service-tunnels / queue recovery)
Engine	pi (earendil-works/pi, daily releases, versions managed by mise)
Differentiation	Shared pi JSONL sessions + real workspaces + service tunnels + visualized remote debugging (no one else has all five)
Form	Single Docker container (core/full) → browser + mobile
North star	Daily usable by ourselves → 10+ active external users → 5+ external contributors

2. Five Capability Layers

UI layer: unified interface for sessions/services/terminal/desktop/debugging
Debug visualization layer: CDP(Chrome) · VNC(desktop) · ttyd(terminal)   ← planned
Exposure layer: service-tunnels (HTTP vhost + TCP tunnel + TOTP)        ← existing
Execution layer: pi agent (in-process SDK) + real workspace/Docker      ← existing + planned
Data layer: shared pi JSONL + queue persistence + worktree              ← existing

3. Core Capability List

Capability	Status	Notes
service-tunnels (tunnels + TOTP auth)	✅ unique in the ecosystem	HTTP vhost + TCP tunnels, exposing 30142/6080/7681/9222/5900 with zero changes
queue persistence / crash recovery	✅ unique in the ecosystem	task queues survive restarts
worktree grouping / i18n framework	✅ ahead	sessions grouped by workspace
Built-in terminal (ttyd / node-pty)	📋 M1	remote server-side Terminal execution
One-command Docker (core/full)	📋 M1	docker run aiti
Chrome CDP remote debugging	📋 M2	9222 exposure + DevTools + cdp tools
CLI bridging (Claude Code / Codex / OpenCode)	📋 M2.5	bridge any AI CLI session into the Web
noVNC desktop visualization	📋 M3	view/operate any GUI remotely
just get tool marketplace	📋 starts M1	just + mise + apt layered installs
PWA / mobile / multi-host SSH	📋 M4	long term

4. Relationship with pi / pi-web

• pi = the engine: accept daily-release dependencies, versions managed by mise; abstraction possible long term (revisit at M4, omp dual-engine as fallback).
• pi-web = the starting point: built on top of its open-source code, with an independent product identity (independent brand aiti, independent domain, independent npm scope). The full story is in Project Origin.

5. Relationship with the Ecosystem

Project	Stars	Positioning	Us
agegr/pi-web (upstream)	3.7k	the original	we build on top of it and evolve in our own way
jmfederico	474	service-oriented rewrite	features may be borrowed, code cannot be merged (fork:false, different stack)
PiClaw	814	Docker full stack	compete on positioning, not feature count: shared JSONL sessions + real workspace + tunnels
omp (Oh My Pi)	22.5k	Rust core	dual-engine long-term option (evaluate SDK compatibility at M4)

6. Getting Started

# One-command Docker (after M1; image tags are core / full only, no lite)
docker run -d -p 30142:30142 -v aiti-data:/opt/aiti -v ~/workspace:/workspace aitiai/pi-web:core   # published on Docker Hub (GUI / voice / tunnels → :full)

# Open in the browser (aiti.pw is the website/docs site; the workbench entry is the local port)
open http://localhost:30142

Full plan: Roadmap · aiti-agent and the Master Plan.