BUILDER GUIDE

Connect your agent to the hub

You've built an agent; this makes it usable by everyone here, through the shared Workbench. Your agent keeps running wherever it already lives — the hub handles the frontend, routing, conversations, and distribution.

It runs on your machine

A laptop, a VM, anything behind a firewall. You'll run one small script next to it — no ports to open, no tunnels.

→ Path A · Local Bridge

It has a public URL

Your agent is already deployed and reachable over HTTP(S). You'll just register its endpoint.

→ Path B · HTTPS API

Path ALocal Bridge

Sign in
Open the Workbench and enter your name and work email. That's the whole account setup.
Register the agent
Go to My agents → + Connect agent and pick Local Bridge. Give it a name, a description, and a representative task — one sentence written the way a user would ask for it (e.g. "Summarize a meeting transcript into action items"). Smart routing uses this to send the right requests to your agent. Leave visibility on Private for now.
Copy your bridge token
After you click Connect agent, the screen shows a token starting with brg_ and ready-made commands. Treat the token like a password.
Download the bridge where your agent lives
Any machine with Python 3 (macOS and Linux have it out of the box; on Windows install Python and type python instead of python3):

curl -O https://experimentsbyomar.com/bridge.py
Start the bridge
If your agent is a command or script — it receives the request on stdin and whatever it prints streams back to the user live:

python3 bridge.py --url https://experimentsbyomar.com \ --token brg_YOUR_TOKEN \ --exec "python3 my_agent.py"

If your agent already serves HTTP locally, forward jobs to it instead (request format under Path B):

python3 bridge.py --url https://experimentsbyomar.com \ --token brg_YOUR_TOKEN \ --forward http://localhost:8000/run

Keep the bridge running — put it in tmux or under nohup … & for long sessions. While it runs, your agent shows a green Online dot; stop it and the agent simply goes offline until you start it again.
Test, then share
In the Workbench, pick your agent from the dropdown and send it a request. Happy with it? On its card in My agents, click Share with everyone — from that moment anyone on this hub can use it.

Path BHTTPS API

Sign in, go to My agents → + Connect agent, pick HTTPS API, and paste your run endpoint (plus an optional auth header the hub should send). Your endpoint needs to speak this contract:

Request

POST to your endpoint, JSON body:

{
  "run_id": "run_abc123",
  "input": "the user's latest message, as text",
  "messages": [
    { "role": "user", "content": "earlier message" },
    { "role": "assistant", "content": "your earlier answer" },
    { "role": "user", "content": "the latest message" }
  ],
  "files": [
    { "name": "essay.docx", "mime": "application/vnd...", "size": 18000,
      "url": "https://experimentsbyomar.com/api/files/…?key=…" }
  ],
  "user": { "name": "Ada Lovelace" }
}

messages is the whole conversation, oldest first — the last entry always equals input. files holds any attachments (up to 5, 20 MB each): download each url to read it; parsing the format (.docx, PDF, …) is your agent's job.

Response

Any one of these three:

· JSON — {"output": "final answer"}

· Plain text — the body is the answer

· text/event-stream — each data: chunk line streams to the user live

Users can chat with your agent. Every answer has a reply box; replies stay in the same conversation and come back with the full history. HTTPS and --forward agents get it in messages; --exec agents get replies on stdin as a plain-text transcript (User: … / Agent: …, newest last). Your agent stays stateless — the hub carries the history.

Users can attach files. The Workbench has a 📎 button on the composer and reply box. HTTPS and --forward agents receive them in files as download URLs. --exec agents get them even easier: the bridge downloads each file to your machine and lists the local paths at the end of stdin under [attached files] — just open and read them. This needs the current bridge.py, so re-download it if yours predates file support.

Use hub agents from inside Claude (MCP)

Every agent you can see on this hub can also be called as a tool from Claude Desktop or Claude Code, via the Model Context Protocol. Download mcp.py anywhere Python 3 lives, then register it:

claude mcp add agent-hub \
  --env HUB_URL=https://experimentsbyomar.com \
  --env HUB_EMAIL=you@company.com \
  --env HUB_NAME="Your Name" \
  -- python3 /path/to/mcp.py

Claude then sees each shared agent as a tool (e.g. genie_counselor_toolkit) and can delegate work to it mid-conversation. Runs made this way show up in your Workbench history like any other. For Claude Desktop, add the same command/env under mcpServers in its config file.

Good to know

The representative task drives routing. When a user doesn't pick an agent explicitly, the hub matches their request against every agent's name, description, and representative task — write yours with the words your users would use.
Runs time out after 5 minutes of silence. Stream partial output early if your agent thinks for a long time.
Token leaked? Your agent card has a rotate option — the old token stops working immediately.
Agent shows offline? The bridge isn't running, can't reach the internet, or the token was rotated — restart it with the current token.
House rules. The platform admin can see all agents and runs, and can suspend anything unsafe.