Case study 02
Sync every coding agent across every Mac.
AGENT-TOOLS keeps skills, commands, sub-agents, and host policy in one git-backed control plane, so Codex, Claude Code, and OpenCode can share the same operating model across machines.
hosts
Claude Code
OpenCode
Codex
agent-tools
AGENT-TOOLS
sync · place · load · verify
Macs
MacBook Pro
Studio Mac
Travel Mac
Fast proof
Turned fragile AI-agent workstation setup into a portable control plane for multi-machine sync, host switching, and token-aware skill loading.
Multi-machine sync
Skills, commands, sub-agents, and AGENTS policy travel through one git-backed source of truth instead of being recreated by hand on each Mac.
Host portability
Codex, Claude Code, and OpenCode receive the right assets in their own runtime homes, reducing the cost of switching agent hosts.
Context economy
The skill-library keeps rare or heavy guidance out of default context while cookbook bundles load the right skills per work mode — a small default context is what keeps multi-host work reviewable.
Problem
Coding-agent productivity collapses when workstation setup drifts. Each host has different folders, command conventions, skill loading behavior, and policy expectations. Without a sync layer, every laptop becomes a custom snowflake.
What it proves
Demonstrates the ability to reduce AI-agent workstation sprawl into one local control plane, then use profiles, skill-library loading, cookbook presets, and drift checks to keep switching cost and token cost under control.
Reliability pattern
Treat agent setup as versioned local infrastructure: one source of truth places host assets, verifies drift, and keeps specialized guidance loaded deliberately.
Sanitized evidence
AGENT-TOOLS is what stops AI-assisted engineering from being a story about one carefully configured laptop. It turns agent setup into a portable workstation layer: skills, commands, sub-agents, and policy live in one source of truth, then each host receives the assets it expects.
Token-aware loading
Not every skill belongs in the default context window. The skill-library keeps specialized guidance available without loading it into every session. That makes the workstation feel broad without making every prompt pay the token cost.
Cookbook work modes
Cookbook bundles are task-mode presets. A frontend cookbook can load UI/UX, GSAP, accessibility, visual QA, and browser verification guidance. A backend cookbook can load API design, migration, security, debugging, and verification guidance. A QA cookbook can load reproduction, regression, and release-readiness guidance. A code review cookbook can load diff risk, test-gap, security-check, and reviewer-note guidance.
What the workstation is not
It is not a team platform. It is a single-engineer reliability system sized for portable AI-assisted delivery across machines and agent hosts.
Privacy notes
No host vendor lock-in is claimed; references to coding-agent hosts use only their public product names. No internal tooling, organization, or repository is named.
View next case study
MCP Router
One shared local MCP router instance for multiple coding-agent hosts, with cached capability lookup, isolated plugin runtimes, and measurable routed-vs-direct overhead.