I built meepr, where AI agents audit the code, file bug tickets, propose features, and implement changes.
The app itself started as standard microblogging 1.0 stuff – auth, profiles, following, likes, "remeeps", DMs, notifications etc.
There is purposefully no search, no recommendations, always chronological, no injected content, no easy way to be found if you don't want to be outside of sharing stuff yourself to your own network, just like the good old days.
On meepr, hashtags act more like Quora topics than post tags. An AI agent called the Tag Agent analyzes posting patterns for each hashtag, while a User Agent builds opinion profiles on everyone based on their meep history – topic affinity, expertise signals, behavior patterns. When you post to a hashtag, the system evaluates whether you belong there. New users get cold start handling. The idea is that in theory #machinelearning should surface people who actually know ML, not whoever's loudest, we'll see. :)
Then there's the self-maintaining part. Every night at 3am, an agent called Steward wakes up, clones the repo into an isolated worktree, runs npm audit and linting, scans PM2 logs for error patterns, and uses Claude with tool use to investigate anything weird. Weekly, it distills everything into themes and generates tickets in YAML format.
Those tickets feed into "Genesis", which runs a full OODA loop. It observes the workspace, orients around the backlog, decides what to work on, then acts – writing code, creating snapshots for rollback, running validation, restarting services if everything passes. Seven "pressure agents" act as internal critics arguing over every proposed change: one obsessed with security, one with performance, one advocating for users, one collecting tech debt, and so on.
Every post also gets analyzed – sentiment, topics, spam scoring, embeddings for semantic search. The V2 memory system builds persistent context about users that feeds back into the gating decisions. Snapshots before every mutation. A separate AI reviews changes before commit. Multiple autopilot modes from "observe only" to "fully autonomous." Secret redaction in all prompts. Major features get approved by me, then ship themselves about once a month.
It's been running for about 4 days now, the agents find real bugs, propose sensible improvements, and the code they generate usually runs, and every time it's managed to unstick itself.
Stack is Node/Express/Postgres/Drizzle with pgvector for embeddings, mix of GPT-5.2 and Sonnet for Genesis and the gating agents, runs on a single DigitalOcean droplet with nine PM2 processes.
I built meepr, where AI agents audit the code, file bug tickets, propose features, and implement changes.
The app itself started as standard microblogging 1.0 stuff – auth, profiles, following, likes, "remeeps", DMs, notifications etc.
There is purposefully no search, no recommendations, always chronological, no injected content, no easy way to be found if you don't want to be outside of sharing stuff yourself to your own network, just like the good old days.
On meepr, hashtags act more like Quora topics than post tags. An AI agent called the Tag Agent analyzes posting patterns for each hashtag, while a User Agent builds opinion profiles on everyone based on their meep history – topic affinity, expertise signals, behavior patterns. When you post to a hashtag, the system evaluates whether you belong there. New users get cold start handling. The idea is that in theory #machinelearning should surface people who actually know ML, not whoever's loudest, we'll see. :)
Then there's the self-maintaining part. Every night at 3am, an agent called Steward wakes up, clones the repo into an isolated worktree, runs npm audit and linting, scans PM2 logs for error patterns, and uses Claude with tool use to investigate anything weird. Weekly, it distills everything into themes and generates tickets in YAML format.
Those tickets feed into "Genesis", which runs a full OODA loop. It observes the workspace, orients around the backlog, decides what to work on, then acts – writing code, creating snapshots for rollback, running validation, restarting services if everything passes. Seven "pressure agents" act as internal critics arguing over every proposed change: one obsessed with security, one with performance, one advocating for users, one collecting tech debt, and so on.
Every post also gets analyzed – sentiment, topics, spam scoring, embeddings for semantic search. The V2 memory system builds persistent context about users that feeds back into the gating decisions. Snapshots before every mutation. A separate AI reviews changes before commit. Multiple autopilot modes from "observe only" to "fully autonomous." Secret redaction in all prompts. Major features get approved by me, then ship themselves about once a month.
It's been running for about 4 days now, the agents find real bugs, propose sensible improvements, and the code they generate usually runs, and every time it's managed to unstick itself.
Stack is Node/Express/Postgres/Drizzle with pgvector for embeddings, mix of GPT-5.2 and Sonnet for Genesis and the gating agents, runs on a single DigitalOcean droplet with nine PM2 processes.
Come meep with me! https://meepr.co/je