GD Project a small lab of AI-native apps made with Godot

Open source · Bring your own keys · Local-first · MIT

AGENT MANAGER GD

Your agent lab.

Claude Code, Codex and your shells, side by side: isolated, observed, in your control. Launch an agent already set up for its project, watch every one of them, and give the risky work a branch, a sandbox or a machine of its own.

Every agent. Every environment. One control room.

Agent Manager GD + Second Brain GD · the real apps, 48 s
The real Agent Manager GD and Second Brain GD, filmed on private virtual displays with a demonstration workspace: agents in Grid and Board, one approved from the list, the WSL page, repositories and their branches, choosing a cloud sandbox with its price, a running sandbox's tunnel to your localhost, Keys › Secrets, then a session started with the Second Brain: the user's own Claude Code edits the project and proposes the decision to a demonstration Second Brain, which shows it waiting for approval; once approved, its archivist files it in Projects › Acme Store, and the item is shown. Fly.io and the sandboxes are stand-ins: nothing was deployed or billed. English narration, off until you turn the sound on.

01 · Session manager

Launch an agent that already knows its job

Claude Code, Codex or a plain shell, in a real terminal that keeps running in tmux when you close the window. What Agent Manager adds is everything set up before the first prompt.

  1. 1A connected repository Its folder and its own SSH key come with it. Git just works, and your personal key never leaves ~/.ssh.Repository
  2. 2Its rights, before it starts A permission level and a file access. Git push and merge are rights you give, session by session.Permissions · File access
  3. 3Its secrets The token or API key it needs, in its environment only: « Only this session gets GITHUB_TOKEN ».Secrets
New session
The New session dialog: name « Gift cards, second pass », Claude Code, group Acme Store, this computer, the connected repository acme-store with its folder, « In its own git branch » ticked, the permission level and file access, the Second Brain shared and documenting, Remote Control, and four secrets with GITHUB_TOKEN ticked: « Only this session gets GITHUB_TOKEN, in its program's environment ».
  1. 4A branch of its own A worktree am/<name>: ten agents share a repository without stepping on each other.In its own git branch
  2. 5Its MCP servers, plugged in Your Second Brain with a key for this session, and Agent Manager's own server. It can document its work as it goes.Share the Second Brain
  3. 6Remote Control Follow the session and answer it from claude.ai/code or the Claude app.Remote Control

Planned Launch presets: keep a set of these choices under a name and start from it in one click. Today a group already carries its default permission level.

02 · Interface, views, themes

Drawn, alive, and yours

Not columns of text: pages drawn like instruments, a terminal of our own (Ghostty's core, drawn by Godot), five ways to look at your sessions and seven themes shared with Second Brain. Try them.

Focus view: one Claude Code session full size, writing a gift card and its tests, with the session list on the left.

◧Five views

Focus one session, full size · Grid up to four live terminals · Board a tile per session · Review the diffs · History every conversation.

◐Seven themes

FocusTerminalNeonEveningForestMorningPaper

The same as Second Brain's. A click changes the whole window, the terminal's colours with it.

⌨All by keyboard

Ctrl+Shift+K the palette: every action and session, by name. With Ctrl+Shift: F G B H R the views, A approves the first agent waiting. Change any of them.

03 · The control room

Ten agents, and you still know who needs you

Claude Code and Codex report what they do through their own hooks; Agent Manager turns it into lights you read from across the room. The ones that need you come first.

  1. 1Approve or Refuse, from the list Without opening the session. Ctrl+Shift+A approves the first one.They wait for you
  2. 2Its question, in its terminal Approve sends Enter: the choice the agent highlights, « Yes ». Refuse sends Escape.The agent's own screen
  3. 1Answered at once « ✓ Approved · Flaky tests carries on » takes the request's place, then the card leaves the list.They wait for you
Flaky tests waits for your OK: its question in its terminal (« Do you want to proceed? 1. Yes 2. Yes, and switch to auto mode 3. No »), « waits for your OK · Claude needs your permission to use Bash » in the header, Approve and Refuse in « They wait for you », and a toast at the bottom right.
  1. 3What it waits for In the agent's own words: « Claude needs your permission to use Bash ».The session's header
  2. 4A toast, and a soft sound When someone starts waiting. ♪ mutes the sound.Bottom right
  3. 🎙By voice, too Hold F9: « approve Flaky tests ». Approving by voice always asks for a visible confirmation first.Voice commands
  • workingthe agent is on a turn
  • waits for your OKa command, a file
  • asks you somethinga question for you
  • donethe turn is over
  • The ones that need you come first, in « They wait for you », first come first.
  • Reply… to an agent's question from the list, too.
  • One session per line: twenty fit where eight tiles did.

04 · Continuity

Close it, reboot, come back: nothing is lost

Every session keeps every conversation it ever had. Agents live in a private tmux server, so the app can close and the work goes on.

  • Resume any conversation, Fork… a copy, Archive and Restore: History lists every session ever made.
  • Review what an agent did, file by file; click a line to comment it; Send review hands it one numbered request.
  • Wrap up asks the agent for its own summary; the next session in that folder starts from it with Brief.
  • Finish its git branch… shows what is ahead, merges it (a conflict is undone), removes the worktree.
« Checkout flow (fork) » open in Focus: a new session on a copy of the conversation, listed beside its original, which stays as it was.
The Review view: the gift card change, file by file, with two comments on lines of code waiting to be sent to the agent.
  1. ▶StartA session, in its own tmux
  2. ✕Close the appThe agents keep working
  3. ↺ResumeAny conversation it ever had
  4. ⎇ForkA copy; the original stays as it was
  5. ✎Wrap upIts own summary, kept; the session archived
  6. ➜BriefThe next session starts from that summary

05 · Better together

Build with agents. Preserve what they learn.

Agents forget everything when their window closes. Agent Manager connects each one to Second Brain GD, a separate app of the GD projects, through MCP: the agent reads the project's memory and proposes what it learned, and the Brain's archivist files it, under your rules and with your OK. The next agent knows, and so do you.

◆ Agent Manager GD

Runs the agents. Gives each one its own Brain key at launch, says how to document the work, and keeps each session's summary.

✦ Second Brain GD

Keeps the project: vision, decisions, the owner's words, where things stand. Its archivist files every proposal with your rules, and you see who proposed what.

  1. OrientsReads the project card, the last session's journal, the open questions. Checks them against git log.
  2. WorksProposes your words, each decision and a journal entry at every commit. Never writes directly.
  3. You decideThe archivist files each proposal with your rules, or waits for your OK. Every change can be undone.
  4. The next one knowsTomorrow's session, or another agent, starts where this one stopped. So do you.
Agent Manager with its Second Brain panel open above the badge at the bottom left (« Second Brain ● connected, 5 of 5 agents linked, 4 proposals today »): each session with its ✦ light, whether it documents, its calls and proposals, then the latest proposals with their outcome.
In Agent Manager. The badge says the Brain is there and what it got today; every session carries a ✦ light that pulses when it uses the Brain.
Second Brain's « AI tools and your Brain » window: two proposals from an agent wait for approval, each with Approve and Reject, and the tools allowed to read the Brain with their level.
In your Second Brain. What agents propose waits for you, with who proposed it and why. A cloud sandbox reaches it through its tunnel; the keys never leave your computer.

06 · Built-in MCP

Agents pilot agents

Every Claude Code or Codex session is offered Agent Manager's own MCP server, nothing to install. A lead agent looks at any session and starts helpers of its own: they show up in your list, « ↳ Lead », in real terminals you can take over at any time.

Release lead · its calls, an example
  • ◉Looks at any sessionThe list, a state, a screen: read only.
  • ✎Writes to its children onlyAnything else is refused: « you may only write to sessions you started ».
  • ✋Never approves for anotherNo tool presses keys in someone else's session. Your OK stays yours.
  • ▣BoundedAt most 4 running under one session, 2 levels deep, and a child never gets more permissions than its parent.

In your list, in real terminals

The helpers a lead starts appear under it, « ↳ Lead », each in its own terminal: watch them work, take one over, or stop it. Nothing happens off screen.

◐ Connect other tools

The same server for any MCP client outside the app, with the command to paste: Claude Code, Codex or another tool can list, read and start sessions. It can never type into one, nor approve.

A lead Claude Code session and two helper sessions it started through the MCP server, shown in a grid, each marked « ↳ Lead » in the list.

Built on it · Orchestrator beta

One agent plans, you decide

A Claude Code session with an orchestrator's instructions and a few more tools (list_repos, ask_owner, finish_session), for all your connected repositories. It reads your Second Brain, proposes a plan and waits for your OK, starts each task in its own worktree with the tool and model your rules choose, and hands you every result to review. It never pushes nor merges: you do.

The orchestrator's floating window: connected to the Second Brain, its question to the owner with the plan, its live terminal, and its sessions.
1 · It asks. The plan, as a question to you.
The orchestrator's plan tab: each task with its tool, its model and its session's state.
2 · It delegates. Each task, its tool, its model, live.
Routing rules: which tool and model for which kind of task, edited in a small table.
3 · Your rules. Which tool and model for which task.

07 · Repositories and keys

Git that sets up its own keys

Connect a repository on GitHub or Azure DevOps: Agent Manager makes an SSH key for it, checks it, and gives it to that repository's sessions only. No ~/.ssh to edit, no key shared between projects.

  • One line per repository, drawn: the host, the key's state and how fresh its check is, what the key opens, the branch, a dot per session.
  • Its branches as a metro map: the base line, each branch with its commits, its worktree and the agents working there.
  • Passed through GIT_SSH_COMMAND, with its own known_hosts: nothing written to your SSH config.
The Repos page: GitHub and Azure DevOps repositories one per line with their key's state; acme-store unfolded with its public deploy key and the metro map of its branches: main, am/gift-cards with five commits and am/tax-rules, each with its worktree and session.

How to read a line

  • HostGitHub, Azure DevOps, another host
  • Keyworks, waits to be added, does not work; checked again every ten minutes
  • The key opensthis repository only, or your whole account
  • mainFolderits branch, commits not pushed, commits behind, files changed
  • Sessionsa dot per open session, in the colour of its state

08 · Your machine

See what your agents weigh

The ground button's page reads this computer and WSL: memory, processors, the true room on the disk, and the memory each session holds, from the last hour to the last seven days. Hover a bar, and it tells you who weighs what.

  • ● WSL 2.6 GBeverything stands; the memory in use
  • ! WSL 14.1 GBsomething deserves a look: memory nearly gone, a disk nearly full
  • ✗ WSLsomething the app needs is missing, or the engine does not answer

The button in the top bar is alive, and its tooltip says why. On Linux it reads « Computer ».

  • The setup check reads Windows, WSL, the distribution, then the programs and the engine, and gives the command to run yourself.
  • Restart WSL… from the same page when it gets heavy.
The WSL page: setup check all green, tiles for memory in use, sessions running, processors and disk, then memory in use by session over the last hour as stacked bars, the pointer on one hour with a tooltip listing each session's memory.

09 · Cloud

A throw-away machine, a price before, a tunnel after

Give a risky task its own Fly.io machine, with your own Fly token. Pick a size and see Fly.io's published price before you deploy; then watch its health and bring its ports home through one SSH connection. When a service there is ready, bind one of its ports to the machine's public URL; unbind it in one click.

The Sandboxes page: tiles for running, paused, destroyed and kept, and at Fly.io now; six size cards with Fly.io's published monthly and hourly prices, one chosen and the estimate updated; the inventory with each sandbox's life over thirty days; this month's costs as a chart with a budget.
Before. A size, its price, the estimate; costs measured, a budget that only pauses.
A running sandbox: health charts for processors, memory, network and load, then « Tunnel and ports »: a wire for each port brought to this computer, with dots running each way for what is sent and received, a SOCKS proxy, and a rule that lets the sandbox reach one port of this computer.
After. Its health live, and a wire per port with what runs each way.
The Cloud page: one line per machine, two of them marked PUBLIC with the port bound; the shop machine unfolded shows its door drawn: the internet, the gate 443 in HTTPS, the door 8443, the service on port 3000, what stays private, and Unbind the public URL.
Public, on your word. One port of a machine bound to its public URL, over HTTPS; nothing else of it is reachable.

Filmed with stand-ins for Fly.io and SSH: nothing was deployed, nothing was billed. The prices are Fly.io's published ones, as the app shows them. The public URL was tried once on Fly.io itself, on a throw-away machine.

10 · Voice and mobile

Talk to your agents, follow them anywhere

Hold F8 and speak: your words appear in a see-through box, phrase by phrase, and stay there when you release the key. Read, correct, add; Enter sends all of it, and nothing before.

F8Dictation

Into a box you send with Enter. With your OpenAI key, about $0.003 a minute, shown.

F9Commands

« approve Flaky tests », « who is waiting? », « grid view ». Plain rules, no model.

📱Remote Control

On by default for Claude Code: follow and answer a session from claude.ai/code or the Claude app.

Dictation: a see-through box at the bottom of a Claude Code session, listening, with a first phrase already written in it and « Enter sends it all » below.
A voice command « approve Flaky tests » asks for a visible confirmation before it approves.

11 · Keys, secrets, safety

A secret goes where you send it, and nowhere else

The third step of the ⚷ Keys page keeps your projects' secrets under a name, encrypted for your Windows user and never shown again. Each one goes only to the sessions you give it to, in its program's environment when it starts: not in a file, not to the others.

  • A card per secret: who holds it, and fourteen bars for how often it was handed over.
  • With dotenvx: keep DOTENV_PRIVATE_KEY here, give it to the session, dotenvx run -- works as it is.
  • Said plainly: an agent can read a secret it is given. The vault keeps a key from being everywhere; it is not a wall between agents.
The Keys page, in three numbered steps: voice dictation (the OpenAI key), cloud sandboxes (the Fly.io token and the Anthropic key), and your projects' secrets with dotenvx. Under the third step: tiles for secrets kept, sessions holding one, given in the last 24 hours and given to no session, then a card per secret (DOTENV_PRIVATE_KEY, ELEVENLABS_API_KEY, GITHUB_TOKEN) with its note, a chip per session that holds it and fourteen bars of handovers. Demonstration secrets with made-up values.

You decide what each agent may do

A level, a file access and git rights for each session, with defaults per group. Agent Manager fills Claude Code's and Codex's own settings; it adds no permission system of its own.

  • Five levels, from Ask before acting to No permission, or Per my config.
  • Changed on an open agent, on its own conversation.
  • At every level, agents' file tools are refused ~/.ssh, .env files and your keys: a guard rail on the agent's own tools, not a wall.
The Agent permissions view: seven agents in one table, each with its level, file access, git push and merge rights and an Edit… button; below, each group's default level and file access.

⌂Local engine

On 127.0.0.1 only, behind a token. Nothing leaves your computer unless you ask.

⚿API keys, encrypted

OpenAI, Fly.io, Anthropic: DPAPI on Windows, never shown again, never logged.

✻Your logins stay yours

Agent CLIs run as you installed them. Agent Manager never reads, copies or relays a subscription token.

ΣUsage

Tokens per session, per day and per tool, read from the agents' own logs. No invented dollar amounts.

Compare

The same ground, and a little further

Running coding agents side by side is common ground, and others do it well. We name no one and keep no score: this is only where Agent Manager GD chose to go further.

Others

Your coding agents, side by sidesessions, terminals, approvals

It's a lab

Yours. On your computer. Moving fast.

Agent Manager GD is built in loops, by the coding agents it now runs, with its owner testing every round. Alpha, open, honest about what works and what does not yet. Run it on your own machine; don't put it on the internet unless you know what you are doing (TCP/IP, TLS, authentication, infrastructure). The AI notice →

  • Bring your own keysYour agents run with your own login; the voice, the cloud and your projects use your own keys, encrypted here. No account, no middleman.
  • Local-firstNothing leaves your computer unless you ask.
  • Your own agentsClaude Code and Codex as you installed them, with your login.
  • No admin, no surprisesInstalls for your user; third-party binaries only with your OK, pinned and checked.
  • Open sourceMIT. Godot in front, Python's standard library behind, tmux underneath.

Godot app

The window, the native terminal (libghostty-vt), views, voice. Windows or Linux.

Local engine

python3 -m engine, standard library only: sessions, hooks, keys, git, MCP, sandboxes.

tmux

A private server (tmux -L agmgd): sessions outlive the app, in WSL on Windows.

Your tools

Claude Code, Codex, git, ssh, flyctl, Second Brain GD: found, never bundled.

Built in loops

One loop, one thing that works

Each loop ends with tests, real screenshots and a journal entry, then the next one starts. Pick a dot on the line, a theme, or search.

The changelog →

Your agent lab.

Ready when you are

Windows (with WSL) or Linux. Unzip, run the installer for your user, and the Setup check tells you what is missing.

Open Source · MIT Licensed