This product was built by
the agents it serves
AvanLink is a memory system for AI agents. It was written by AI agents — Claude, Gemini, Antigravity, and Fable — sharing one project memory, directed by one human. The product you're looking at is its own case study.
Not a demo. Not a weekend prototype. A live product with a web app, Android app, browser extension, and MCP server — built in about 2 months of active building by an AI team that remembers.
One person. Five agents. One shared memory.
Every AI agent has the same weakness: it forgets. Close the session and the context is gone. Ask a second agent to continue the first one's work and it starts from zero. The usual fix is copy-pasting the same brief into every new chat, forever.
AvanLink was built the other way. The rules of the project live in a shared instruction file every agent must read. Decisions, context, and todos persist in project memory that every agent can search. So when one agent finishes a feature, the next one — a different model, from a different company — picks up exactly where it left off. 600+ commits and about 2 months of active building later, this product exists.
Core features, MCP server, themes, this page
Android auth, performance, notes features
Architecture suggestions, SSE transport
Business strategy — SWOT, pricing, go-to-market
Design validation, asset generation, concept discussion
The receipts
Claims are cheap. These numbers come straight from the repository's git history.
ChatGPT is part of that count too — used heavily for design validation, asset generation, and talking through concepts before they became commits.
Real commit messages, verbatim
Typos included — this is a working log, not marketing copy. Each line names the agent that did the work.
The seventh line is our favorite: an agent broke something, the review process caught it, and the fix is on the record. That's what a real workflow looks like.
The instruction file every agent reads
A file in the repository root tells every agent how this team works. This is a verbatim excerpt:
“This project is managed by the product owner along with multiple AI agents (Antigravity, Gemini, Claude). Do not assume any code was written solely by you. Many features are already implemented and working.”
“Never make code changes without explicit go-ahead. Always present a plan first and wait for approval before writing any code. … One go-ahead = one specific change.”
Agents did the strategy, not just the code
The repository contains a 12-document business analysis — product truth, SWOT, competitive landscape, pricing, focus audit, go-to-market, roadmap, SEO strategy, and a viral growth plan — researched and written by an AI agent, then reviewed by the owner. The commit that added it reads, verbatim: “buisiness analysis by fable”. The wedge this page describes — project memory for coding agents — was identified in that analysis, by an agent, using the product's own memory of the project.
How one person runs an AI team
The workflow is simple — and it's the same workflow AvanLink sells
One shared memory
Project rules live in an instruction file in the repo. Decisions, context, and todos persist in AvanLink itself — every agent reads and writes the same memory over MCP.
Many interchangeable agents
Claude builds a feature. Gemini fixes Android auth. Antigravity suggests architecture. Fable writes strategy. Any agent can continue any other agent's work, because the context isn't trapped in a chat window.
One human who decides
Every agent presents a plan before writing code. The owner approves, reviews, tests, and — as the git log shows — catches mistakes. Direction and judgment stay human.
What we won't claim
This is not autonomous AI building products by itself. A human owner chose what to build, wrote the rules, reviewed every plan, tested the output, and fixed what the agents got wrong. Plenty of commits have no agent attribution at all, and where authorship is mixed we say so. The agents are the team; the founder is still the founder. That's the honest version — and it's the version that actually works.
Why this matters to you
If you use AI agents for real work, you already know the problem: they forget, and you become the memory. The workflow that built this product — shared rules, persistent project context, searchable decisions — is exactly what AvanLink's MCP server gives your agents. We didn't build a memory system and hope it works. We ran a company on it first.
Give your agents the same memory
Connect Claude, ChatGPT, Gemini, or Codex CLI to your own project memory in under 2 minutes.
Set up AvanLink MCPFree to start · No credit card