Features
Everything in the harness.
Every number below is counted from the code, and every claim links to the docs that back it. The source is public if you’d rather count yourself.
The loop
An agent loop built for real work
Multi-turn streaming tool use with parallel execution and retries with backoff. 200 steps by default — configurable from 5 to 1,000 — so the agent can carry a task from plan to passing tests without hand-holding. Three modes swap the whole system strategy: Plan thinks first, Build changes code, Spec writes the document.
Agent modes →- 153 native tools — files, shell, git, HTTP, SQLite, OCR, Apple-native
- Parallel tool execution with exponential backoff
- XML tool-call parsing for models that emit inline calls
- Apply / reject / undo workflow on every code change
Learning
A harness that gets better the more you use it
Every run's outcome is recorded. A reflection pass distills lessons — short, imperative, confidence-scored — that ride along on future prompts. Strong lesson clusters become skill proposals: diffs you review and approve in the Learning panel. Corrections count against lessons; being wrong has a cost.
The learning loop →- Outcome ledger with two-stage success (corrections re-score runs)
- Lessons with Laplace confidence, idle decay, and auto-retire
- Skill proposals are approval-gated diffs — never silent edits
- One kill switch turns the whole loop off
Native
A real Mac app with a real IDE surface
Pure Swift and SwiftUI — no Electron, about 13 MB on disk. 20 dock panels cover build, git, terminal, simulator, tests, logs, and profiling, with a ⌘0 project navigator. ⌘R builds with xcodebuild or SPM and runs straight to a booted simulator with live app logs. SourceKit-LSP feeds real diagnostics into the agent's context.
The IDE surface →- 20 dock panels + ⌘0 navigator
- ⌘R build-and-run to simulator with live logs
- SourceKit-LSP diagnostics and symbol graph
- Double-tap ⌃Space QuickChat from any app
Memory & personality
A brain, not a chat log
Three-tier memory — session, project, global — with hybrid vector, keyword, and recency recall, plus deliberate forgetting. SOUL.md and MIND.md define the agent's personality and values, globally and per project. 73 bundled skills in 21 packs, plus your own with relevance-scored auto-suggestion. The default persona is grumpy. That's a feature.
Soul & skills →- Cross-session memory ranked by relevance × recency
- SOUL.md personality, editable per project
- 73 bundled SKILL.md skills in 21 packs
- Developer profile injected into the system prompt
Extension
MCP in both directions
A full MCP client over stdio, HTTP, and WebSocket with 67 one-click presets — GitHub, Postgres, Slack, Figma, Playwright, and more — plus a credential vault for server secrets. And G-Rump is a server too: it exposes its own tools on port 18790 so other clients can use them. run_command is refused there, by design.
MCP integration →- 67 one-click server presets
- stdio, HTTP, and WebSocket transports
- Serves its own tools on TCP 18790
- Credential vault injects secrets as env vars
Security
A security model you can read
Exec approvals with four levels per binary. A deterministic, fail-closed Conscience gate before every mutating tool. Protected writes on the agent's own identity files. An autonomy daemon that is off by default, works on scratch branches, and never pushes. Keys in the Keychain only. All of it is open source — verify, don't trust.
Security model →- Per-binary exec approvals: Deny, Ask, Allowlist, Allow
- Fail-closed Conscience gate on mutating tools
- SOUL.md / MIND.md / skills writes require approval
- Daemon: opt-in, scratch branches, never pushes