AI Integration · In Development
AgentDex
A git-native registry for AI agents. It versions agents, syncs them across machines, and consolidates divergent learnings back into one canonical file.
Client
Personal Project
Year
2025
Duration
Phase 1
Role
Solo Developer
What it does
AgentDex treats an AI agent's learned context like code. The registry is just a git repo, where each agent is a markdown file and machines push their local learnings to peer branches. A deterministic merge pass unions identical sections, then an LLM reconciles genuinely divergent ones and flags contradictions. Think git plus a package registry plus a merge strategy for what an agent has learned.
The problem
You run a specialist agent on multiple machines. Each copy learns different things. They diverge. AgentDex fixes this. It versions agents, syncs them across machines, and consolidates divergent learnings back into one canonical agent. Then it redistributes it.
How it works
One repo is the registry. registry.json is the git-native catalog. Each agent is a single markdown file at agents/<name>.md. The body is the learning store, structured by markdown ## and ### sections. main is canonical. Each machine pushes learnings to its own peer/<machine-id> branch.
Consolidation is deterministic first. Identical sections pass through untouched. Only genuinely divergent sections hit the LLM. The LLM merge pass reconciles each divergent section and flags real contradictions instead of silently overwriting. The result commits to main, the version bumps, and a <name>@v<version> git tag is cut.
The CLI loop is clean. Run agentdex init, agentdex add, and agentdex register to push main and branch to peer/<machine>. When an agent learns something new, run agentdex push to commit and push the peer branch. Run agentdex consolidate --all to union and merge into a new canonical version with a tag. Then agentdex pull adopts canonical main everywhere.
Consolidation uses the Anthropic API when ANTHROPIC_API_KEY is set. A --deep flag routes to Opus. If no key is present, it falls back to a fully deterministic offline merge with zero API calls.
The stack is TypeScript, Node >=20, and a pnpm monorepo. It uses zod, gray-matter, commander, and vitest. The Anthropic SDK is an optional peer dependency.
Status
Phase 1 is shipped. The core loop is implemented as a CLI and engine. Unit and E2E tests pass. Everything past that is roadmap. The Next.js dashboard, Supabase run history, contradiction resolution UI, custom git merge driver, and server-triggered consolidation are not built. This is a CLI and core engine. No live UI. No hosted service.
Built with
The build log
Development updates
No development updates yet. Check back soon for progress updates and insights.
