Gentle-AI configures the AI coding agent you already use with persistent memory, a structured workflow, and review evidence. It never installs an agent itself: it writes native configuration for 17 agents, including Pi, OpenCode, Claude Code, Codex, Cursor, VS Code Copilot, Gemini CLI, Kilo Code, Kimi Code, Qwen Code, Hermes, Windsurf, and OpenClaw, and snapshots each config file before every write. Because each integration uses that agent's own capabilities, features such as delegation and review differ between agents.
Engram, its memory layer, has the agent record decisions and lessons as it works and consult them before asking you, so context accumulates across restarts and context compaction. Organic Driven Development (ODD) keeps understood, small changes lightweight and gives substantial work a single recoverable feature document. It applies test-first development by default where a runnable test and a clear expected outcome exist: a failing test per requested rule, then the implementation, then refactoring with tests green.
Receipt-Driven Development (RDD), on by default, freezes the exact change before review so the evidence belongs to that version. A read-only risk assessment then picks the review depth, from a structural readback to a four-lens review for high-risk changes. The gentle-ai binary, not the model, reads change state from disk and returns the only valid next review transition, so different agents, models, and sessions reach the same step.
Gentle-AI also bundles a skills library loaded by task, an optional Context7 MCP server for live library documentation, a read-only code symbol graph, a deny-list that blocks SSH keys and credential files, per-role model assignment where the agent supports it, and optional personas and themes. Gentle Shell is a separate package with its own Pi runtime, agents, and interface.
Features
- Engram memory: the agent saves decisions and lessons and checks memory before asking you again
- Organic Driven Development: proportional workflow that keeps small changes light and gives larger work a recoverable feature document
- Test-first by default: one failing test per requested rule before implementation, plus regression tests for touched behavior
- Receipt-Driven Development: reviews run against a frozen change at a depth chosen by risk
- Deterministic transitions: a binary, not the model, decides the next valid review step from on-disk state
- 17 agent integrations: native configuration for Pi, OpenCode, Claude Code, Codex, Cursor, Gemini CLI, and others
- Skills library: skills load automatically when a task matches
- Context7 MCP: optional live framework and library documentation
- Security deny-list: blocks SSH keys, environment files, and credential files
- Config backups: every agent config is snapshotted before each write
- Health check: a read-only doctor command reports installation health
- Model assignment: configure agent and review-role models where the agent supports it
