A desktop workspace for the coding agents you already run — editor, diff viewer, terminal, browser and a task board, one topic at a time.
Agents run inside Topics instead of one process each, so 200 sessions answering at once fit in 162 MB of RAM.
One topic per thing you are doing. Each holds a chat, a file tree and code editor, a Git view with diffs, a terminal and a browser. Switching topic switches all of them at once, so you stop rebuilding context every time you change subject.
The browser is part of the tool, not a link that opens elsewhere. It runs inside the topic, an agent can drive it, and you watch the real DOM rather than a video of it: text stays sharp and selectable, and it costs a fraction of a pixel stream.
Close the lid and come back. Sessions survive a restart with their whole scrollback, not just the last screen. Most embedded terminals lose everything that scrolled past while you were away, which is exactly what you need when an agent worked for twenty minutes without you.
Go back to before it went wrong. Long agent sessions rarely fail with an error; they take a wrong turn that is obvious only later, after the conversation has built on it. A checkpoint returns the session and the working tree to a saved point, so undoing is not an argument with the agent about what it did.
Hand work to a board instead of babysitting it. Describe a task, and an agent picks it up with everything it needs: working directory, model, effort level, and its own git worktree. Several agents can edit the same repository at once because each has a real checkout on its own branch.
Approving, landing and publishing are three buttons. Accepting an agent's work is not the same as merging it into main, which is not the same as pushing it. Collapsing them into one is how code nobody read ends up on a remote.
Your agents, your keys. It drives the claude-code and codex CLIs you
already pay for, or talks to the APIs with your own key. No credential of ours
in the middle, no proxy: your prompts go straight to the model vendor.
Grab the latest release:
.dmg for macOS, .exe or .msi for Windows, .deb or .rpm for Linux.
Updates install on restart, after you confirm.
First launch on macOS. Builds are not notarized yet, so macOS blocks them. Go to System Settings → Privacy & Security → Open Anyway. Once only.
Topics runs a small server on your machine. Your topics live there, which is why the app works with the network unplugged. Your phone on the same Wi-Fi can reach it after a one-time six-character approval.
Getting in from outside your network needs a relay, and a relay is a machine somebody has to run.
| Plan | What you get |
|---|---|
| Free forever, no account | Everything local, plus your home network. Unlimited topics, projects and agents. |
| Subscription | Reachability from anywhere, and seats for your team. |
The licence is verified offline, so a billing outage can never downgrade the machine in front of you. An expired or unreadable token falls back to the full free plan, never to a locked app.
A session is an array of messages, not a process:
| Sessions at once | All answered | Whole server | Wall clock |
|---|---|---|---|
| 8 | ✓ | ~2 s | |
| 64 | ✓ | ~4 s | |
| 200 | ✓ | 162 MB | 5.6 s |
At equal counts a CLI session costs 432 MB against 2.3 MB native: 190x.
An IDE gives you one project per window, and the next window costs nearly as much as the first. Same three repositories, same machine:
| Each extra project | |
|---|---|
| Cursor | +889 to +1039 MB |
| VS Code | +261 MB |
| Topics | +0.07 MB |
Empty, Topics is 164 MB on disk against 1.3 GB for Cursor. Loaded with 22 projects and ~1000 topics it sits at 440–745 MB.
Method, the runs that contradicted my earlier claims, and what none of this
proves: bench/README.md.
Your conversations stay on your machine. No analytics, no crash reporting. The paid relay is the one thing that routes traffic through our infrastructure, and only if you turn it on — PRIVACY.md has the detail.
Don't put Topics on the public internet. Behind Tailscale or a tunnel, add your own authentication. Vulnerabilities: SECURITY.md.
Build from source
Requires Bun and Node.js 20+.
git clone https://github.com/armonia/topics-app.git
cd topics-app
bun install --frozen-lockfile
(cd client && bun install --frozen-lockfile)
bun run build:client # client → public/
bun run start # http://localhost:3333public/ is a build artefact and nothing rebuilds it on its own: after a change
under client/src you run the build again. While you are working on the client,
use bun run dev:client (Vite HMR) instead of rebuilding. Why it is not
watched: docs/build-watch-decision.md.
The desktop shell needs the Rust toolchain and embeds
public/ at compile time, so build the client first:
cd desktop-tauri/src-tauri && cargo run # dev
cd desktop-tauri && cargo tauri build # installersCI builds the official installers from tauri-vX.Y.Z tags. The pre-v2 Electron
shell lives on the electron-archive branch. Dev workflow:
CONTRIBUTING.md.
Environment variables
Nothing here is required: provider, model and keys are in Settings, and
Settings wins over the environment. These exist for headless runs, CI and
containers. Copy .env.example to .env.
| Variable | Description | Default |
|---|---|---|
PORT |
Local server port | 3333 |
APP_DATA_DIR |
Where conversations and app data live | ~/.openclaw |
SERVER_HOST |
127.0.0.1 keeps the server on this machine only |
all interfaces |
ANTHROPIC_API_KEY |
Only for the direct claude provider, not the CLI |
— |
OPENAI_API_KEY |
Only for the direct openai provider |
— |
ELEVENLABS_API_KEY |
Text-to-speech and Scribe v2 dictation | — |
MOONDREAM_API_KEY |
Browser vision grounding | — |
AI_PROVIDER |
Pins one provider | auto |
CLAUDE_MODEL · OPENAI_MODEL |
Model ids | — |
GATEWAY_URL · GATEWAY_TOKEN |
OpenClaw gateway | http://127.0.0.1:18789 |
STT_PROVIDER · STT_LANGUAGE |
Dictation engine and language | auto |
Picking the default provider. With AI_PROVIDER unset your keys decide:
ANTHROPIC_API_KEY → claude, else OPENAI_API_KEY → openai, else
GATEWAY_URL → openclaw, else claude. That choice holds as long as it stays
connected.
The order claude-code → codex → claude → openai → openclaw only picks
the replacement once the current default goes offline, so with an API key set
the CLIs above it never get a turn.
A pin beats connectivity. A pinned provider that is offline stays the target and its chats never answer, which is deliberate: a pin is an instruction.
STT_PROVIDER=auto tries ElevenLabs Scribe v2 → OpenAI gpt-transcribe →
Deepgram Nova-3 → Groq Whisper turbo → local whisper.cpp, using whichever keys
you have.
MIT licensed, provided "as is" without warranty of any kind.
Topics is an independent project by Armonia, not affiliated with or endorsed by Anthropic, OpenClaw or ElevenLabs. It talks to third-party services with keys you provide, under each provider's own terms. See THIRD-PARTY-NOTICES.md.
MIT © Armonia