Giving Claude, Cursor, and 13 Others a Shared Brain

What it does

Gentle-AI is an ecosystem configurator, not another agent installer. It layers persistent memory, spec-driven development workflows, curated skills, MCP servers, and a teaching-oriented persona onto 15 existing agents—from Claude Code and Cursor to Kimi Code and OpenCode. The goal is to stop agents from behaving like stateless chatbots and start treating them as context-aware teammates that know your stack and your past decisions.

The interesting bit

The project treats agent chaos as a workflow problem. It defines delegation triggers—such as touching two non-trivial files or reading four-plus source files—to force a context switch or review, preventing long monolithic sessions from spiraling. For OpenCode, you can even route different SDD phases to different models, using a cheap model for exploration and a premium one for architecture.

Key highlights

• Supports 15 agents with tailored integration depths, from full sub-agent orchestration to solo-agent harnesses.
• Engram provides cross-session persistent memory with a TUI and CLI for searching past decisions and consolidating project context.
• Automatic compressed, deduplicated backups of agent configs, auto-pruned to five snapshots with optional TUI pinning.
• Per-phase model profiles for OpenCode via gentle-orchestrator, allowing cost and capability optimization by workflow step.
• gentle-ai doctor performs read-only health checks of tool binaries, state, memory reachability, and disk space.

Caveats

• Gemini CLI support is explicitly marked experimental in the agent matrix.
• The delegation triggers are described as expected conventions rather than hard runtime enforcements, so discipline is still required.

Verdict

Try it if you are tired of re-explaining your codebase to agents every session. Ignore it if you prefer lightweight, unopinionated AI assistance without workflow guardrails.