What is Gentle-AI?
Gentle-AI is a single Go binary that rewrites the configuration of coding agents you already have, such as Claude Code, Codex, Cursor, OpenCode, Gemini CLI and Pi. A terminal installer asks which agents, components and persona you want, then writes prompt files, skills, MCP server entries, permission rules and hooks into each agent's own config. It never installs an agent itself.
The components are Engram (persistent memory served over MCP from a local SQLite store), a skills library, the Context7 documentation server, a permissions deny-list, a persona, the GGA provider switcher and a theme. On top sit two workflows: Organic Driven Development (ODD), which keeps small changes inline and only writes a feature document for work that spans sessions, and Receipt-Driven Development (RDD), which freezes a finished change and reviews it at a depth set by its risk. Alan Buscaglia (Gentleman Programming) maintains it, funded by GitHub Sponsors and consulting.
Is Gentle-AI worth it?
Gentle-AI is less a skills pack than a configuration manager: it owns parts of every supported agent's settings and keeps them in sync, which is both its use and its main risk. It has about 9,100 GitHub stars and 896 forks, and it published 297 releases between 28 February and 7 November 2026, when v4.4.0 shipped. That is more than one release a day.
The workflow argues against heavy planning. ODD sizes a task by whether it can be resumed from the diff, not by file count, and the orchestrator only spawns a subagent when it can name one of four reasons: exploration, parallel writers, a context backstop or an independent verifier. The project's own benchmark, one run per cell, says a 3-line bug fell from 335 seconds and $0.339 to 77 seconds and $0.105 once routing ran it inline. The same notes admit a large task still took 17.5 to 17.9 minutes against 6.1 minutes in plain Codex CLI. The earlier spec-driven workflow (SDD) was retired in v4.0.0, so older tutorials no longer match.
What it puts on your machine deserves a plain list. The installer script adds one line to your shell startup file for PATH. Engram stores memories in local SQLite and can export them to a .engram/ folder for git or replicate to an Engram Cloud server you point it at. GGA installs a second global binary, and the optional CodeGraph tool installs its own CLI. On Claude Code it adds startup, review and telemetry hooks. Every install, sync and upgrade first snapshots the configs it will touch and keeps the 5 most recent backups.
The permissions component is a real guard with stated holes. It denies agent reads and edits of ~/.ssh, .env files, cloud credentials and key files on Claude Code, OpenCode and Kilo Code, and on OpenCode it blocks force pushes and recursive rm on the home directory. The docs say reading a secret through bash or grep is not blocked, and that wrappers such as env or xargs slip past the wildcard rules. Telemetry is on by default: a random install ID, version, OS, installed agents and review counters, sent at most once a day, plus model names and token counts from supported agent hooks. DO_NOT_TRACK, GENTLE_AI_TELEMETRY=0 or gentle-ai telemetry disable turns it off, and nothing is sent on the first run.
Against Superpowers, which is a fixed sequence of skills you drop into one agent, Gentle-AI is heavier to adopt and lighter on small tasks. It earns its place where a team mixes agents and wants one memory and one review standard across them; for a single Claude Code user, it is a lot of moving configuration to inherit.
Where it earns its place
A freelance developer who switches between Claude Code, Codex and Cursor during the week
Stop re-explaining the same project decisions every time a session restarts
Engram is wired into all three as an MCP server, so a decision saved in one agent is searchable from the others. Memories stay in local SQLite; engram sync writes them to .engram/ if you want them in the repo. No cost beyond model usage.
A tech lead at a five-person startup whose developers each prompt their own way
Give every developer's agent the same guardrails and review depth
Each developer runs the installer with the same preset; the permissions component denies reads of ~/.ssh and .env files, and RDD sizes review by risk instead of running four reviewers on every change. Expect to re-run gentle-ai sync often, given the release pace.
A backend engineer fixing small bugs with an agent that over-plans
Get one-function fixes done without a planning pipeline
ODD keeps understood, low-risk work inline with one failing test written first. The maintainers' n=1 benchmark shows a 3-line bug going from 56 turns to 10; treat that as an indication, not a measurement of your codebase.
Before you commit
- Setup
- brew install or a curl-to-bash script on macOS and Linux; on Windows, go install with Go 1.25.10 or later. Run gentle-ai to pick agents, components and persona, then gentle-ai doctor for a read-only health check. Release signatures can be verified.
- Integrations
- The README counts 17 integrations, including Claude Code, Codex, Cursor, OpenCode, Kilo Code, Gemini CLI, VS Code Copilot, Windsurf, Antigravity, Kiro, Qwen Code, Hermes and Pi (through Gentle Shell); the agents doc also lists GitHub Copilot CLI. Conductor is detected but gets no managed writes. Context7 and CodeGraph are added as MCP servers.
- Data and privacy
- No hosted service is required: memory is local SQLite, with optional export to git or an Engram Cloud server you configure. Anonymous telemetry is on by default and documented field by field; it excludes paths, repo names, prompts and code but includes public model IDs and token counts. GitHub lists no security advisories.
- Cost at scale
- Free under MIT. The real cost is model usage, which the routing rules try to reduce, and the time to re-sync configs as releases land; the maintainer sells consulting for team rollouts at prices not published.
- Limits
- 366 open issues. Capabilities vary by agent, for example Cursor and Antigravity do not get the managed review agents. The docs track main and are dense with migration detail, and the permission rules cannot parse shell wrappers or nested shells.
What’s good
- One installer configures 17 agents with the same memory, skills and review rules
- Small tasks run inline without subagents or planning documents
- Config snapshot before every write, with restore and a read-only doctor command
- Telemetry is documented field by field, with a preview command and several opt-outs
What’s not
- Close to 300 releases in about eight months; behaviour and file layouts keep moving
- Telemetry and RDD review are both on by default
- Permission deny-list does not stop reading secrets through bash or grep
- Per-agent features differ, and Pi needs the separate Gentle Shell package
Best for
Developers or small teams running two or more coding agents who want shared project memory and one review standard across them.
Gentle-AI alternatives
Gentle-AI: common questions
- Is Gentle-AI worth it?
- A thoughtful, well-guarded setup layer for people who run several coding agents, held back by release churn and documentation few will read to the end. We rate it 4.0 out of 5.
- How much does Gentle-AI cost?
- Free and open source (MIT). You pay for the model usage of the agents it configures.
- What are the downsides of Gentle-AI?
- Close to 300 releases in about eight months; behaviour and file layouts keep moving; telemetry and RDD review are both on by default; permission deny-list does not stop reading secrets through bash or grep; per-agent features differ, and Pi needs the separate Gentle Shell package.
- Who is Gentle-AI best for?
- Developers or small teams running two or more coding agents who want shared project memory and one review standard across them.
- How much does Gentle-AI cost for a team or in production?
- Free under MIT. The real cost is model usage, which the routing rules try to reduce, and the time to re-sync configs as releases land; the maintainer sells consulting for team rollouts at prices not published.
- Is Gentle-AI safe for company data?
- No hosted service is required: memory is local SQLite, with optional export to git or an Engram Cloud server you configure. Anonymous telemetry is on by default and documented field by field; it excludes paths, repo names, prompts and code but includes public model IDs and token counts. GitHub lists no security advisories.
- What are the best Gentle-AI alternatives?
- The closest alternatives we have reviewed are Superpowers, Everything Claude Code (ECC), BMAD Method, claude-mem and Spec Kit.
Answers checked Nov 13, 2026.
Make Gentle-AI? Show this rating on your site
