Quickstart: ChatGPT Workspace Agents

Workspace Agents is ChatGPT’s builder for repeatable-task agents — you describe the agent in natural language, connect it to tools, then share it with your workspace, run it in Slack, on a schedule, or trigger it through an API. It’s one of OpenAI’s two recommended migration paths off Agent Builder, which shuts down November 30, 2026 (the other path is the code-first Agents SDK).

Workspace agents speak MCP, so one connector gives every agent in your workspace access to 4,921 live-data tools across 1,300 sources.

Requirements: a ChatGPT Business, Enterprise, or Edu workspace. Custom MCP connectors are added by a workspace owner/admin (or a member they’ve authorized as a developer).

① Enable custom MCP connectors (admin, once per workspace)

  1. Go to Workspace Settings → Permissions & Roles.
  2. Under Connected Data, enable Developer mode / Create custom MCP connectors.

② Add Pipeworx as a connector

  1. In ChatGPT: Settings → Connectors → Create (custom connector).

  2. Fill in:

    FieldValue
    MCP server URLhttps://gateway.pipeworx.io/pipeworx-catalog/mcp
    Namepipeworx
    AuthenticationNone

Pipeworx is anonymous by default — no account or key needed. Append ?apikey=YOUR_KEY to the URL later for higher rate limits (free signup).

③ Build the agent

  1. Create a new workspace agent and describe its job in the instructions.
  2. Under the agent’s tools/apps, add the pipeworx connector you created.
  3. Add a routing rule to the agent’s instructions so it prefers live data over its training set:
When the user asks for current or factual data — economic stats, SEC/EDGAR
filings, stock or crypto prices, weather, FDA drug info, patents, procurement
notices, or government records — call the pipeworx tools (ask_pipeworx)
instead of answering from memory. Prefer structured Pipeworx data with citations.

④ Test

Ask the agent:

Using Pipeworx, what was the US unemployment rate last month?

You should see a tool call to ask_pipeworx and a cited BLS number in the reply. Then publish the agent to your workspace, wire it into Slack, or put it on a schedule — every copy shares the same connector.

Migrating from Agent Builder

If your Agent Builder workflow used Pipeworx as an MCP tool, the connector setup above is the direct equivalent — same URL, same Authentication: None, same routing rule. What changes:

  • Workflow graphs don’t transfer. OpenAI’s export produces Agents SDK code, not a workspace agent — for natural-language agents, rebuild the agent’s job description from your workflow’s instructions and re-test.
  • Tool selection moves from the Agent node’s tool picker to the workspace connector + the agent’s tools list.
  • Scoped connections still work: swap the URL’s scope segment (e.g. https://gateway.pipeworx.io/trade/mcp for only the trade packs). Browse scopes in the directory.

Building in code instead? The Agents SDK quickstart covers the HostedMCPTool path.

Troubleshooting

Can’t find developer mode / custom connectors. It’s workspace-plan-gated (Business/Enterprise/Edu) and off by default — a workspace owner has to enable it under Permissions & Roles.

The agent answers from memory. Strengthen the routing rule in step ③; name ask_pipeworx explicitly.

Rate limited. Anonymous tier is per-IP and shared by everyone behind your workspace’s egress. Add a free API key to the connector URL.

Last reviewed July 14, 2026