Local agents

Two ways to use it

• In your browser: switch the sidebar to Code. The AI writes Python and runs it in a sandbox inside your browser — nothing is installed and nothing touches your files.
• On your computer: get an agent key from the owner and point your coding agent at the API below. This is the full agent experience: it can read and edit your project and run commands, with your approval.

Connection settings

• Base URL: https://titan-ai.me/v1
• API key: your agent key (starts with pa-) — do not share it
• Model: auto (Titan AI picks the best working provider), or an exact model ID from GET /v1/models
• Streaming and tool calling are supported, so agents can run and report back as they work

Quick setup

Cline / Roo Code (VS Code): Settings → API Provider → OpenAI Compatible; Base URL https://titan-ai.me/v1; API Key: your agent key; Model: auto.

Continue: add a model with provider: openai, apiBase: https://titan-ai.me/v1, your key, and model auto.

aider:

aider --openai-api-base https://titan-ai.me/v1 \
      --openai-api-key pa-YOUR-KEY \
      --model auto

Cursor: Settings → Models → OpenAI API Key: your agent key; enable "Override OpenAI Base URL" and set https://titan-ai.me/v1; add a custom model named auto.

Use it from the website (one line, then one click)

The quickest way: on the site switch to Code mode and press Set up. A window shows one command — copy it, paste it into your Terminal, press Enter, and come back. The page connects by itself.

curl -fsSL https://titan-ai.me/go | bash

That script downloads the agent to ~/.titan-agent/, starts the bridge in the background with a pairing code it keeps in ~/.titan-agent/pair-code (so later runs stay paired), makes a working folder ~/titan-agent-workspace, and opens this page with the pairing code in the URL so everything connects automatically.

Then just ask in the chat, for example 帮我打开计算器. Every action appears as a card with Run / Skip — nothing runs without your click. Stop the bridge with pkill -f 'titan-agent.py --serve'.

Manual alternative, if you prefer to see every step: download titan-agent.py, run python3 titan-agent.py --serve, then click Connect and paste the code printed in the terminal. Use --workdir ~/some/folder to choose where file tools may act.

Titan Agent in the terminal

The same single file also works on its own in the terminal: it can open apps, run commands and edit files, asking you before every action.

1. Download titan-agent.py and save it anywhere.
2. Get an agent key from the owner.
3. Run it:

export TITAN_API_KEY=pa-YOUR-KEY
python3 titan-agent.py                      # local server
python3 titan-agent.py --base-url https://titan-ai.me/v1
python3 titan-agent.py --workdir ~/demo     # files stay in this folder

Then just type what you want, for example 帮我打开计算器 or 在这个文件夹里做一个 todo 网页. Every action is printed and needs a yes before it runs. --yes skips the prompts (only for requests you trust), and --allow-outside lets file tools leave the working directory. File tools stay inside the working directory; run_shell can run any command you approve.

Controlling apps without screenshots (AX)

Before reaching for GUI mode, the agent can ask the frontmost app for its controls: ui_tree lists every element as index | role | title | position | size, and ui_action presses an element or types into a text field by its index. This works with any model — no vision needed — and it is much cheaper and steadier than clicking pixels.

• It needs macOS Accessibility permission for the terminal running the bridge (same as GUI mode).
• It always acts on the frontmost app, so the agent should open or focus the app first (its open_app tool does that).
• Quality depends on the app: Safari/Chrome pages, System Settings and many native apps expose labels ("Reload this page", "Example Domain"), while a few (for example Calculator) expose positions but no names — for those, GUI mode or AppleScript is the way.

GUI mode (experimental)

With the bridge connected you can also tick GUI mode (experimental, Gemini) under the composer. The agent then works like a person: it takes a screenshot, sees it with Gemini, clicks or types one action, and looks again — up to 8 steps per request.

• macOS only for now, and it needs three permissions for the app you started the bridge from (usually Terminal): System Settings → Privacy & Security → Accessibility, Screen Recording and Input Monitoring, then quit and reopen that app and restart the bridge. A quick way to check: ask the agent to call gui_check. If screenshots show only your wallpaper with no windows, Screen Recording is the missing one.
• The model only sees screenshots of your screen (sent to Gemini). Coordinates it sends are in the screenshot's image space; the bridge scales them back to your real screen.
• Every single click, key and scroll is shown as a card and waits for Run. Nothing happens without your click, so watch the cards before approving.
• Prefer the normal (non-GUI) tools when they can do the job: they are faster, cheaper and easier to check. GUI mode is for things that only exist in a window.

Limits

• Agent keys: 300 requests/day and 30/minute by default (agents make many small calls per task).
• Classmate keys: 30/day and 5/minute.
• All keys share the site's daily pool, so please keep long agent runs reasonable.

Safety

• The agent runs on your machine and asks you before it acts. Titan AI never touches your computer or your files.
• Read what the agent proposes before you approve it — especially commands that delete or upload files.
• Keep prompts school-appropriate, and remember that prompts are stored so the owner can keep the service healthy.

No agent yet? The browser Agent mode works with no setup, and the API docs have plain-JSON examples if you want to build your own small agent.