Roadmap & status
Grumpy enough to tell you what’s broken.
G-Rump is early software. Most product pages hide that; this one is the map. Three lists, stated plainly: what you can rely on, what’s shipped but rough, and what’s coming. If the rough column annoys you enough to fix something, the good first issues are waiting.
Works today
Use it for this now.
The native app itself
Pure SwiftUI, 13 MB, macOS 14+. Onboarding, Welcome window, Settings, 20 dock panels, ⌘0 navigator — the surfaces are solid.
Multi-provider BYOK
Anthropic, OpenAI, Google, OpenRouter with native wire formats and streaming tool calls. Keys validate on save and live in the Keychain only.
Provider docs →The tool system
160 native tools with parallel execution, retries, and exec approvals on every shell command. This is the oldest, most-tested part of the harness.
Tool docs →MCP, both directions
Client over stdio/HTTP/WebSocket with 67 presets; server on TCP 18790 that refuses run_command. Battle-tested against real servers.
MCP docs →The security model
Exec approvals, the fail-closed Conscience gate, protected identity-file writes, an off-by-default daemon. Deterministic code, heavily tested.
Security model →Skills, soul, memory substrate
73 bundled skills, SOUL.md/MIND.md personality, three-tier memory store with ranked recall. The plumbing works; see the next column for the payoff.
Soul & skills →Rough — being ironed out
Real, shipped, and imperfect.
Chat polish
The core loop streams fine, but the conversation experience has rough edges — rendering hiccups, occasional awkward states. Being sanded down release by release.
File what you hit →Agentic coding reliability
Some tasks land clean end-to-end; some wander and need you to steer. This is the honest state of the whole category — we're just the ones saying it on the website.
Build & run to simulator
New in 2.1. Works on the happy path (xcodebuild/SPM → booted simulator → live logs); environment differences will surprise it. Bug reports here are gold.
Report an edge case →The learning loop
All the mechanics shipped and tested — outcome ledger, lessons, reflection, approval-gated skill proposals. Whether it compounds into a meaningfully smarter agent over weeks is the experiment, and it's running in the open.
How it works →Local Ollama models
New in 2.1: keyless, live-discovered, offline-friendly. Known issue: paywalled :cloud models list but fail on use.
The known issue →Next
In order. No dates.
GRumpKit
Extract the harness into a SwiftPM library so you can build your own surface on it instead of forking the app. Boundary design is the first step — and it's open for discussion.
Join the design →Notarized builds
Current builds are ad-hoc signed, so macOS quarantines them on first launch. The signing pipeline exists; a Developer ID is the missing piece. Until then, the download page tells you exactly what to expect.
Install notes →Homebrew cask
brew install --cask g-rump. Tap-hosted first, homebrew/cask once builds are notarized.
The issue →Sparkle auto-updates
The update service is already wired into the app; it needs an appcast feed and signing keys to come alive.
The issue →Deeper debugging
Breakpoint-level debugging and on-device runs, building on the 2.1 build engine.
Progress lands in the changelog; discussion happens on GitHub.