loopctl
- State
- Live
- Kind
- Open source
- Role
- Solo founder, engineer, and open-source maintainer — thesis, architecture, implementation, and releases.
An agent-native project-state store for AI development loops.
loopctl is the shared memory and the rulebook for AI coding agents: what each one is working on, who signed it off, what it spent, a record of what it did that nobody can edit afterwards, and a searchable store of what the team has already worked out.
Problem
AI coding agents are unreliable narrators of their own work: they approve themselves, call unfinished work done, and lose hard-won decisions between sessions. Making them trustworthy means putting the state, and the rules, outside the agent.
Approach
Assumed the agent will try to route around any rule it can read, and put the rules underneath it where it cannot: separation between teams enforced by the database itself, separate keys so no identity can both do a piece of work and sign it off, short-lived permissions on the steps that matter, and a history that can be shown not to have been edited after the fact.
Role
Solo founder, engineer, and open-source maintainer — thesis, architecture, implementation, and releases.
Stack
- Elixir
- Phoenix
- PostgreSQL
- Row-Level Security
- Oban
- Cloak
- OpenAPI
- escript CLI
Outcome
- A piece of work cannot be reported or signed off by whoever did it — the rule is enforced by the server, not by convention
- One team's data is kept from another's by the database itself, not by a filter a caller could skip
- The history can be shown not to have been edited after the fact
- Agents read and add to one shared store of what the team has learned instead of working it out again every session
Arrival
Want work like this shipped for you?
Every project starts with a short discovery conversation — no obligation, no boilerplate pitch. Tell me what you're building.