ApeironCode — AI Coding Agent Platform

Context & Goals

ApeironCode is an open-source platform for running AI coding agents that you can actually trust in a real repo. The design centers on control: agents propose a plan, pass through gates, execute in a sandbox, and leave a complete audit trail — with the ability to roll back at any checkpoint.

Architecture

• Planning gates: agents must produce a plan that is approved (manually or by policy) before any mutating action runs.
• Sandboxed execution: tool calls run inside an isolated environment with explicit permission policies governing file, network, and shell access.
• Checkpoints & rollback: every step is snapshotted so a run can be reverted cleanly when an agent goes off-track.
• Audit logs: each tool invocation is recorded with inputs, outputs, and policy decisions.

Highlights

• Built a typed tool-calling layer so agents interact with the system through well-defined, permissioned actions.
• Terminal-first UX with a React control surface for inspecting plans, diffs, and run history.
• Policy engine that turns "what's allowed" into declarative rules rather than ad-hoc prompt instructions.

What I'd do next

• Expand the policy library with org-level presets.
• Add multi-agent collaboration with shared checkpoints.