Small Mac host
David is measured locally as an 8.4 MB app and 3.8 MB DMG, because the host is Swift-native and the heavier work is delegated.
Comparison
David is the native shepherd layer for AI work.
OpenClaw, Hermes, Claude Code, Codex, and VS Code are strong at different jobs. David's bet is narrower: keep AI workers, memory, worktrees, approvals, and loops together in a small Mac app that normal people can supervise.
The simple story
Other tools run agents. David keeps the whole work loop legible.
Built-in loop surface
Loops, workers, worktrees, PRs, CI, CodeGraph, memory, and open gates are first-class product surfaces.
File-bus continuity
David's file-bus gives agents shared chat, ledgers, next actions, events, and layered workspace memory that humans can inspect.
Needs You gates
Consequential actions move into approvals, so the worker team can continue without pretending every step is safe.
As you scroll
The comparison is not only features. It is where the user lives.
Frameworks begin with configuration.
OpenClaw and Hermes are powerful systems for agents, channels, model providers, skills, schedulers, and gateways. That power is real, but the user often starts in setup mode.
CLIs begin with a terminal.
Claude Code and Codex are excellent coding workers. David is designed to supervise them: keep context, route approvals, remember work, and show the loop state after the terminal scrollback is gone.
Editors begin with developer muscle memory.
The old David VS Code extension proved the control-plane idea. The native app removes the editor dependency and makes the shepherd model the product, not a panel inside another product.
David begins with the goal.
The first user model is simple: Goal, Plan, Workers, Needs You, Done. The deeper machinery remains available, but the product starts from supervision rather than agent plumbing.
Measured footprint
The host should be small when the work is already heavy.
These numbers were measured on our Mac on . They compare local installed or packaged footprint, so they are directional rather than a universal benchmark.
Ease of use
A practical score for supervised multi-worker loops.
This score is our product-fit rubric, not an industry standard. One point each: single install, no terminal required for the primary path, visible loop dashboard, durable memory surface, and built-in approval handoff.
Direct Mac app, visible loops, file-bus memory, Needs You gates, and a beginner-friendly work model.
Easy to start as a coding agent, but the loop shepherd layer is not the primary product model.
Strong terminal worker with a clean install path; supervision still lives mostly in prompts, terminal state, and user discipline.
Onboarding helps, but daemon, workspace, provider, and channel setup make it more of an agent system than a native Mac control room.
Powerful learning-loop agent with many surfaces, but setup and provider/runtime choices are more technical.
A great editor, but supervised AI worker loops are not native to the editor itself.
Feature chart
Where David is different.
| Tool | Primary shape | Native loop integration | Memory | Approvals | Best at |
|---|---|---|---|---|---|
| David native app | Swift-native Mac shepherd layer | Built-in loops, continuity graph, worktrees, PRs, CI, CodeGraph, and open gates | File-bus with chat, ledgers, events, next actions, and layered workspace memory | Needs You gates are part of the product | Supervising AI workers without living in terminal or editor plumbing |
| Old David VS Code extension | VS Code control plane extension | Agent tree, jobs, dream cycle, file-bus, and approvals inside VS Code | Shared file-bus, ledgers, chat, and long-term memory files | Needs You panel in VS Code | Developer-first proof that the shepherd model works |
| OpenClaw | Open-source personal assistant and gateway | Agent loop is core to the runtime, with gateway and session routing | Markdown memory files plus memory search, active memory, and dreaming options | Tool policy and sandbox controls, not a David-style Mac approval surface | Always-on personal automation through messaging channels |
| Hermes Agent | Self-improving agent platform | Closed learning loop, skills, schedulers, gateway, and terminal backends | Agent-curated memory, skill creation, session search, and user modeling | Power-user controls through CLI, gateway, and configuration | Research-grade, self-improving agents that can run beyond one laptop |
| Claude Code | Terminal coding worker | Excellent for autonomous coding turns, subagents, git, and shell work | Project context and session state inside the coding tool, not David's shared file-bus | Terminal/tool permission flow | Doing focused coding tasks inside a repo |
| Codex | OpenAI coding agent app and CLI | Local coding agent, app workflows, worktrees, automations, and git review surfaces | Codex config and project/thread context, not David's shepherd memory model | App and CLI review/permission surfaces | Running OpenAI coding agents locally and reviewing their work |
| VS Code | Electron-based code editor | Extensible editor, not an AI worker loop system by default | Project files, extensions, and editor state | Extension-dependent | Editing, debugging, extensions, and developer workflows |
Sources and notes
What the comparison is based on.
Early access
Use the workers. Let David shepherd the loop.
David is built to sit above the coding agents you already trust and make the work visible, resumable, and governable.