OpenClaw Product Benchmark Notes¶
Date: 2026-05-06
Scope¶
This benchmark looked at OpenClaw as a product experience, not as a feature-by-feature clone target. The comparison focused on installation, first run, onboarding, diagnostics, and trust cues that make a developer tool feel polished.
What Was Tried¶
- Cloned
https://github.com/openclaw/openclaw.gitinto a temporary checkout. - Read the OpenClaw README, package metadata, launcher, install docs, getting-started guide, onboarding docs, and CLI references.
- Ran the source launcher with
node openclaw.mjs --help. - Ran the published package with
npm exec --yes openclaw@latest -- --help. - Ran
npm exec --yes openclaw@latest -- onboard --help. - Ran isolated non-interactive onboarding with temporary
OPENCLAW_STATE_DIRandOPENCLAW_CONFIG_PATH. - Ran isolated
openclaw doctor --non-interactiveagainst that temporary state. - Compared DevCD with
devcd --help,devcd onboard --help,devcd smoke --json,devcd doctor --json, anddevcd onboard --preview --no-tui.
Observations¶
- OpenClaw's published package path is excellent:
npm exec --yes openclaw@latest -- --helpworks immediately and prints version, command map, examples, and docs. - OpenClaw's source checkout failure is also useful: it explains that
dist/entry.(m)jsis missing, givespnpm install && pnpm build, and recommends installing the built package for releases. - OpenClaw's getting-started path is a concrete success chain: install, onboard, verify gateway status, open dashboard, send the first message.
- OpenClaw's onboarding is very powerful but can overwhelm: the help output exposes many providers and flags at once.
- OpenClaw's failed non-interactive onboarding still classifies the failure, prints the target URL, and gives exact remediation commands.
- OpenClaw's successful
--skip-healthonboarding prints state writes, workspace status, gateway defaults, and JSON. - OpenClaw Doctor feels product-grade because it names each subsystem, explains why it matters, and offers exact repair commands.
- OpenClaw Doctor also has rough edges: on Windows it emitted symlink errors and attempted completion repair during a non-interactive diagnostic run.
- DevCD already has a quieter local-first proof:
devcd smoke --jsonis daemonless, deterministic, and does not require credentials. - DevCD's weakest product gap is not capability; it is first-run guidance. The user needs one obvious command that explains what will happen, what success looks like, and what to run next.
- DevCD's
onboard --previewis close, but showingAgent-ready workspace - skippedabove an Agent-Layer proposal feels contradictory.
Improvement Checklist¶
- [x] Add a zero-write
devcd welcomecommand that acts as the product front door. - [x] Make
devcd welcome --jsonusable by automation and docs tests. - [x] Update top-level help so new users start with
devcd welcome, thendevcd onboard --preview. - [x] Make
devcd smokeend with the next onboarding command, not just pass/fail checks. - [x] Remove the confusing
Agent-ready workspace - skippedline from preview flows that already show an Agent-Layer proposal. - [x] Document the benchmark and connect it to the install/getting-started path.
- [x] Verify with focused CLI tests, docs build, and
make check.
Product Principle¶
DevCD should beat OpenClaw on the local-first continuity use case by being calmer, safer, and faster to prove: one install check, one preview, one explicit apply, one Action Packet. It should borrow OpenClaw's concrete success chain and diagnostics without copying its gateway breadth or noisy side effects.