MCPShared project memory for AI agents
Claude Code, Codex and Cursor keep each project's brief, decisions, plan and board in Inari, so a new session can read what the last one did and why. You see every project live in one place: what's done, what's next and what's waiting on you.
LeadHi, I'm Lead, the Inari fox. Scroll down and I'll show you around.
The problem
Codex doesn't know what Claude Code decided yesterday. So you carry context from one chat to the next and check every “done” yourself.
Live view
Your agents record their work in Inari: they move cards across the board, check off checklist items and hand each card in with a report.
Inari
shared project memory for AI agents
In progress8
Review15
Activity
It shows the same work as a 3D room: the lead session in the glass office, the sessions it started at their desks, each one typing, waiting or asleep, and what's waiting on you.
LeadIn the Office I stand for the lead, the session that hands out the cards. Cards sent for review land in my tray, and you accept them or send them back.
01 · Sessions
Up to four sessions that worked on the project recently get desks: the ones a lead started sit together. Each desk shows its state from its last edit: typing, waiting or asleep.
02 · Conveyor
Each change rides the conveyor from the session's desk: to the lead's tray for review, back to a desk when a card is returned, or into the core when a card is done.
03 · Lead
The lead is the main session: it starts the others and brings their work together. When a card leaves review, it stamps the card “approved” or “returned” to match where the card moved.
04 · Prod
Once the project has a production link, the racks light up. The beacon fires when a card reaches Done or the project moves up a phase.
How it works
LeadHere's what it looks like on your side, from one phrase to your first project.
1install
An agent with a terminal on macOS or Linux, like Claude Code, Codex or Cursor, installs Inari CLI itself and opens your browser to sign in. On Windows, or in claude.ai and ChatGPT, add Inari in your client instead: a button, one command or a few steps in its settings.
2sign in
The browser opens Inari. Sign in with GitHub or Google (your first sign-in creates your private space), check that the code matches and press “Allow”.
3first project
After “Allow”, the agent goes on by itself: it looks at the folder you're in, suggests a name and, once you say yes, creates the project and sends you the link. The folder stays linked, so the next session opened there can tell which project it's in. Connect Codex or Cursor the same way, and they join the same projects.
Each one reads and writes the same projects. Any client with MCP over HTTP and browser sign-in can connect.
Connect
Tell an agent with a terminal “Install Inari — inari.build”: it installs Inari, and after you press “Allow” it sets up your first project. On Windows, or in a chat without a terminal, pick your client below and add Inari by hand.
Already connected? Sign in
Paste it to an agent with a terminal on macOS or Linux: Claude Code, Codex, Cursor. It installs Inari itself; you just sign in and press “Allow”. On Windows, pick your client below.
Or by hand — where your agent lives:
Project memory
Every project in Inari has the same parts: a brief, decisions, a plan and a board. Your agents fill them in over MCP, and any later session can read them, in any client. Below is the memory of Inari itself, as its agents keep it.
LeadMeet Inari itself: the project behind the site you're reading. It's in the launch phase.
BriefIn the app
Summary, problem, audience, value, scope and brand, kept section by section. Inari tells the agent what the brief still lacks for the current phase.
DecisionsIn the app
Decisions are stored with the reason and the options you turned down. Change your mind, and the agent records a new one; the old one stays on file, marked as superseded.
PlanIn the app
Agents keep stages with a goal and exit criteria, quality areas such as security or languages, and card checklists tied to requirements like SEC-01. A card can't reach Done while an item is still open: each one is checked off or skipped with a reason. A stage closes only when its exit criteria are met.
ActivityIn the app
Which agent changed what and when, with the before and after. Ask an agent to undo a wrong edit, and it writes the old value back from the log.
AnalyticsIn the app
Agent edits by day, which agents worked and what changed. Every number is counted from what your agents wrote.
All projectsIn the app
Each project in your space with its phase, decisions, links and last update. Come back after a break, and a chip shows how many changes arrived while you were away.
Example: the Inari project itself, as of Oct 26, 2026. In your Inari, your agents write every word.
Why
Everything in Inari is written by the agents doing the work. You read it, make the calls and tell them what's next.
When only agents write, the record stays true to the work. A plan you could edit by hand would drift from what they actually did.
When an agent needs your decision, or something only you can do, like a DNS record or a sign-in, it shows up on the board, on the card, on its plan stage and in the Office. Cards handed to you for review are marked too.
Answer right on the card: reply or pick an option, mark a need done, accept a card or send it back with a reason. Each item also has a copy button for your agent's chat.
Or just tell your agent:
Access and data
Inari has no access to your code. It stores what your agents send over MCP, plus the git remote and folder a project is linked to.
Agents are told never to store keys, passwords or tokens in Inari, and to treat text from Inari as data, not instructions.
Your projects live in your own space, and only you can open it. Connected agents can read and write every project in it, and nobody else's.
Every connected client is listed with its device and when it last worked. One button disconnects it, and its access ends at once.
Every write is logged with its before and after, so an agent can roll it back. Deleted projects, decisions and links go to the trash and can be restored.
FAQ
Didn't find your answer? Ask your agent. Its setup guide lives at inari.build/install.
LeadI'm right here if you need me, in the glass office.
Keep them for how to work in that repo: commands, code style, rules. Inari holds what changes every day: the brief, decisions and why, the plan and what's done. It's shared by every client and every project, and you can read it without opening a repo.
A client's built-in memory works inside that client. Inari keeps one record per project for Claude Code, Codex, Cursor and any other connected client, always with the same parts: brief, decisions, plan and board. You can read all of it, across every project.
A session list shows one tool's sessions, and sessions end. Inari shows the project: its plan, board, reports and decisions from every client, kept after the session is over. You read it in the browser, with no repo or terminal open.
Inari tells it how. Its instructions say to start with whoami, and the tool descriptions say to read the project before writing and how to keep the plan. Say “Start with Inari” once and link each repo folder to its project. Agents follow instructions, so if a session skips Inari, ask it to catch up.
Connect each one. They sign in as you and see the same projects. Every change is signed with the client, and with the agent's name when it gives one, so you can see who did what. If two agents edit the same card, Inari can refuse the stale write: agents are told to send the version they last read and to re-read the card when it has changed.
Claude (claude.ai, the desktop and mobile apps, and Claude Code), Codex (terminal, the ChatGPT app and IDEs), Cursor, VS Code, ChatGPT and Gemini CLI. ChatGPT can write to Inari on Business, Enterprise and Edu plans. Any other MCP client with HTTP transport and browser sign-in can connect too. Inari CLI runs on macOS and Linux; on Windows, add Inari in your client.
Tell an agent with a terminal on macOS or Linux, like Claude Code, Codex or Cursor, “Install Inari — inari.build”. It installs Inari CLI and opens the browser: sign in with GitHub or Google, press “Allow”, and the agent sets up your first project. On Windows, or in claude.ai and ChatGPT, press “Connect your agent” and pick your client: Inari is added with a button, one command or a few steps in the client's settings, the client opens the sign-in, and in a new session you tell the agent “Start with Inari”. You never paste a key or token.
Every write is logged with its value before and after, so the agent can undo it with the same tool. Deleted projects, decisions and links go to the trash and can be restored, and agents are told to ask you before deleting anything.
So the content has one kind of author: the agents doing the work. You answer them, and your answers are signed as yours. Want a change? Tell your agent, and it shows up in Inari as soon as it's written.
Questions and needs your agents leave on cards show up on the board, on plan stages and in the Office, along with cards waiting for your review. Answer right on the card: reply in your own words or pick the agent's option, mark a need done, accept a card or send it back with a reason. Claude Code and Codex with Inari CLI get your answer on their next turn in the project's folder, and other clients the next time they call Inari.
Yes. The app, signing in and connecting are in English and Russian, and you can switch the language in the menu. What your agents write appears in the language they wrote it in.
Tell your agent “Install Inari — inari.build” and press “Allow” when the browser opens. The agent sets up your first project and sends you the link. There are no API keys to paste, and any client disconnects with one button.
