A local control plane for agent work that outlives the chat session

What it does

LoopX sits between you and your coding agent as a durable local control plane. It stores the objective, current todos, evidence logs, human approval gates, and compute quota in a compact state kernel so that multi-day or multi-agent work stays reviewable and restartable across turns and tools. The agent runtime still writes the code; LoopX just makes sure the loop remembers what it was doing.

The interesting bit

The project treats control state as an “agent-native Kanban board” where moves like claim, gate, and writeback are validated operators rather than free-form chat history. It also enforces a quota and scheduler hints so your agent doesn’t burn API tokens on useless transitions while you are away.

Key highlights

• Agent-agnostic bridges for Codex App, Codex CLI, Claude Code, Cursor, and custom shell runners.
• Durable local-first state for goals, todos, evidence, and concrete human judgment gates.
• Quota-aware ticks and safe fallbacks to prevent runaway compute.
• Peer-agent support with claims and leases—no single leader identity required.
• Runs fully offline with no runtime dependencies outside Python’s standard library.

Caveats

• The project self-labels as “loop agents early” status, so expect rough edges.
• It explicitly refuses autonomous production control—dangerous permissions and final ownership stay with the human.
• The README only mentions macOS and Linux shells; Windows support is unclear.

Verdict

Worth a look if you are trying to orchestrate multi-day research, benchmark, or issue-fix workflows across multiple agent runtimes. Skip it if you just need a single-turn code generator or fully autonomous deployment.