Claude Plan Comparison

Generated: 2026-05-24

Summary

Claude's plan and Bob's plan agree on the big direction: build a vendor-neutral agentic startup operating system using gstack-style workflow, gbrain-style memory, graph-based code/project context, runtime competition, and closed-loop self-improvement.

The biggest difference:

Bob's plan originally emphasized OpenClaw as the live gateway and recommended building a control/memory/workflow layer above it.
Claude's plan emphasizes gbrain as memory source of truth, Claude Code + gstack as cockpit, Paperclip as always-on company manager, Hermes as background self-improving runtime, and Graphify for larger codebases.

After comparison, the stronger combined plan is:

OpenClaw: gateway/channel interface, not foundational OS.
Paperclip or Mission Control-like control plane: company/org/task/governance layer.
gbrain/Second Brain v2: source-of-truth memory.
gstack: workflow discipline.
Graphify/Graphiti: graph-backed project/code intelligence.
Hermes: self-improving background worker runtime.
Claude Code/Codex: engineering workbench/runtime.

Where Claude's Plan Is Stronger

It gives Paperclip a more central role as the company/governance layer.
It is more concrete about self-improvement loops: telemetry -> evals -> skill-author PRs -> approval -> improved skills.
It frames gbrain as git/markdown source of truth, which is more portable than a black-box memory DB.
It correctly scopes Graphify to larger repos instead of installing it everywhere.
It gives a practical weekly operating picture.

Where Bob's Plan Is Stronger

It is more skeptical of migration risk and unverifiable hype claims.
It treats OpenClaw as a proven live gateway because Henry already uses Discord/Telegram workflows.
It emphasizes project isolation more explicitly: repo, env, Supabase, task queue, memory namespace, permissions.
It separates architecture from tools: control plane, context router, memory router, task router, runtime router, policy engine, evaluator.
It avoids replacing Second Brain blindly before a schema/retrieval comparison.

Questionable Claims In Claude's Plan

gstack/gbrain star and benchmark claims should be treated as directional until verified from primary sources.
Paperclip capabilities are promising, but many sources are marketing posts or early community reports. It should be piloted, not trusted blindly.
"Skip OpenClaw" is too aggressive. Discord-first always-on routing still matters for Henry's workflow, even if OpenClaw is no longer the foundational OS.
Claude Code as cockpit is strong for desk work, but weak as the always-on human/company interface.

Updated Recommendation

Build "WiderWings Agentic OS" with these layers:

Gateway: OpenClaw initially.
Company control plane: Paperclip evaluated against Mission Control.
Memory: gbrain-style markdown/git + Second Brain v2.
Workflow: gstack skills and artifacts.
Code/project graph: Graphify pilot on MedSchools.ai when repo size justifies it.
Runtime workers: Claude Code, Codex, Hermes, OpenClaw sessions.
Evals/self-improvement: telemetry and skill PR loop before autonomous promotion.

Sources Checked

gbrain README: https://github.com/garrytan/gbrain/blob/master/README.md
gstack: https://github.com/garrytan/gstack
Graphify Claude integration: https://graphify.net/graphify-claude-code-integration.html
Graphify token reduction discussion: https://techjupjup.com/en/ai/graphify-guide/
Paperclip GitHub org: https://github.com/paperclipai
Paperclip guide: https://paperclip.ceo/blog/building-a-zero-human-company-a-quick-start-guide
Claude Code hooks: https://code.claude.com/docs/en/hooks-guide