Install
https://github.com/Gentleman-Programming/gentle-ai/blob/main/docs/quickstart.mdWho It's For
You already use an AI coding agent, yet the same thing keeps happening: once it finishes writing code it wipes the context and starts the next round from scratch; it has no stable grasp of your project; and after the work is done, the only way to prove what it actually changed is to read the diff yourself, line by line.
Gentle-AI targets exactly this kind of everyday work. It does not replace your agent; instead it layers a deterministic engineering environment on top of the one you already have, which the official docs summarize as "The deterministic engineering environment". Its focus comes down to three things: memory, workflow, and evidence.
It suits developers who have a real repository and want their agent to remember project conventions over the long run; team leads who want the agent to advance along a fixed process instead of improvising every time; and situations that need a verifiable record for every change, so collaboration and handover stay easy.
Put the other way, it does not address requests like switching to a better editor or model: it governs how your existing agent works, not which tools you use. If your need is a one-off script and you don't need a traceable record, think first about whether adding another layer of configuration is worth it.
Your First Attempt in 3 Minutes
Follow the single official path: open the Quickstart.
https://github.com/Gentleman-Programming/gentle-ai/blob/main/docs/quickstart.md
Prerequisites: you already have a supported agent installed (Claude Code, for example), and it is allowed to read and write your project directory. Gentle-AI does not ship an agent of its own; it only writes configuration into the agent you already have. The official docs claim coverage of 17 integrations, and the per-agent differences are recorded in docs/agents.md — that comparison table also tells you which capabilities a given agent can use.
Before you start, spend a minute reading that comparison table and confirm that the agent you use daily is listed. If it isn't, switch to a listed agent before discussing configuration — trying to work around it only gets you an environment that won't install and whose effect you can't explain.
The install location is the target directory named in the section of the Quickstart for the agent you use. The homepage only says it performs native configuration and doesn't list each agent's on-disk path, so don't guess directories from memory and don't copy the path from another agent.
Once it's installed, don't let it make big changes right away. Start with a read-only job: read through the project structure, explain how it intends to understand the repository, and require a plan before it touches anything. This step writes no files and runs no commands, so the risk is lowest. Whenever writing files, executing shell commands, or making external requests is involved, have it list the files it intends to change and the actions it intends to run, item by item, and only let it proceed after you confirm.
There are two observable success signals: open a different session and the agent can still state the conventions that already exist in your project, which shows the memory was written to disk; and within a single session it follows a fixed process — plan first, then act, then leave a record you can check — instead of just handing back an answer. The official description of the full path is: one prompt carries through to a commit with review results.
Use Cases
Case one: one prompt takes you from an idea to a commit with review results. The section on the repository homepage demonstrates exactly this line: memory retrieves the project background, workflow constrains the steps, evidence explains the changes. The evidence stops there — the homepage shows the chain but has no step-by-step interface details, so the real sequence and artifacts should be confirmed against the Quickstart and docs/intended-usage.md.
Case two: add memory, workflow, and checkable evidence to your agent. The approach is to write configuration for the chosen agent as the Quickstart describes; the officially listed options include persistent memory, curated skills, MCP servers, and personas. Note that the homepage gives only the names of these configuration items — it doesn't say which files map to which agent, nor which one wins when they conflict; for that, go back to docs/agents.md and each agent's own documentation.
What the two cases share is this: you don't have to replace the agent you already have; all you gain is the ability to remember and to explain. The homepage gives no quantified data on efficiency or quality gains, so don't expect it to prove the benefit for you.
Limits of Use
It is a layer of configuration, not a new agent or IDE: you must first have a supported agent for the configuration to land in. The official docs claim 17 integrations, and the homepage names only some of them; for whether a given editor or command-line tool is on the list, check the comparison table in docs/agents.md rather than assuming everything works.
The homepage describes the second validation only as optional and limited in scope, without giving thresholds or the situations it applies to, so don't treat it as a quality gate — important changes still need your own review.
Details like install locations and where each configuration item lands are scattered across the individual agents' documentation; the homepage only provides the entry point, so when you copy paths, check which agent's section you're reading.
The documentation, the official site, and the wiki are currently all in English; copy commands and paths exactly as they are and mind the casing. The repository homepage also links to both the official site and the wiki, so for configuration questions about a specific agent you can look there first for an existing answer.