Releases: mtecnic/clusterspace
v1.1.0 — Autonomous AI
ClusterSpace is a tiled desktop workspace for terminals, embedded browsers, and SSH+tmux sessions — with an optional AI co-pilot that can drive any pane, verify visually, and pursue goals autonomously until they're done.
This release lands the autonomous AI stack. Set a goal, walk away, come back when it's verified done.
🔗 Repo · Wiki · Quick Start · Goal Runner overview
What's new since v1.0.0
The headline: a Goal Runner that pursues an objective until verified — the AI can't "just stop." It either calls claim_complete (and verification has to pass), or abort_with_report (graceful give-up). Otherwise the loop runs.
Shipped across 9 phases:
| Tool registry + plugin hot-reload | 52 tools dispatch through a unified registry. Drop a .js file into your config dir to add your own — hot-reloaded without restart. |
| Paged tool results | read_terminal_output, browser_get_content, browser_get_axtree return cursor-paginated envelopes. Large outputs no longer blow the context window. |
| Per-pane conversation isolation | Builder agent in pane A doesn't pollute Tester agent in pane B. Keyed by (provider × workspace × pane). |
| Goal-scoped action policy | Risk ladder: read_only → write_local → network_get → network_write → spends_money. Plus allowlist, denylist, sandbox dir. Tools beyond ceiling prompt the user. |
| Persistent goal log | Every step checkpointed to disk. Resumable in the dashboard. 50-goal cap, 500-step cap per goal with head+tail preservation. |
| GoalRunner engine | Server-side loop with wall-clock cap, four verification types (shell exit-code, model YES/NO judge, manual, predicate), agent-status tracking. |
| Critic + replan loop | A second model judges progress every N steps. Verdicts (PROGRESSING / STUCK / ACHIEVED / MISLED) inject guidance or short-circuit to verification. Optional cheap provider for cost-efficient oversight. |
| Goal observability dashboard | `Ctrl+Shift+G`. Live step log with color-coded tools, critic rail showing the AI's internal negotiation, abort / delete / prune. Status-bar pill pulses while goals run. |
| Vision-grounded browser verification | `browser_verify_visual_state` (strict YES/NO vision judge) and `browser_describe_screen` (free-form). Catches modal overlays, error toasts, visual regressions that DOM checks miss. |
Plus: full per-tool reference, 6 default personas (Builder, Reviewer, Tester, Admin, Monitor, Claude Code Expert), 2 skills, 2 task templates, all overridable from your config dir.
Try it
```bash
git clone https://github.com/mtecnic/clusterspace.git
cd clusterspace
npm install
npm run rebuild # native modules (node-pty + Electron)
npm run dev
```
Configure a provider (Claude, OpenAI, Ollama, LM Studio — anything OpenAI-compatible). Open the goal dashboard. Give a pane a goal like "Install nginx, verify with `curl -sf http://localhost\`." Watch it run. Walk away.
See Installation for OS-specific notes (Windows needs VS Build Tools, macOS needs Xcode CLT) and AI Providers for setup walkthroughs of each provider.
A full wiki
44 pages covering everything from keyboard shortcuts to goal-runner internals to writing your own plugin tools. Start at the wiki Home.
What it isn't
- A SaaS — everything runs on your machine, your data goes only to the AI provider you choose
- A replacement for an IDE — pairs well with one
- A mobile app
- Free of bugs — please open an issue if you find any
Compatibility
- Windows / macOS / Linux
- Node 20+ for building from source
- Tmux on remote hosts (only if you want SSH session persistence)
- Any OpenAI-compatible AI provider (or skip the AI entirely)
What's next
See the Roadmap. Highlights: light theme, goal pause/resume from UI, JSON-predicate evaluator, browser recipe recorder, test suite, customizable keybindings. PRs welcome — see Contributing.
MIT licensed. Built with Electron + React + TypeScript.