Paid workspaces · Claude Code + Codex + OpenClaw + plain HTTPS

Your support agent, inside your coding harness

Point Claude Code, Codex, or OpenClaw at helpGPT and ask your docs anything — answers come back grounded in your content with citations, or the agent says it doesn't know. No MCP client? Plain HTTPS works from any script.

Why MCP

Ask your docs without leaving the harness

The agent answers from your ingested help content with [n] citations — or says it doesn't know. Same guard, caps, and history as the widget.

Every text-based free tool, inline

Markdown converters, sitemap generate/validate/extract, link extraction, transcript analysis, ROI, signatures, PDF ask/summarize — results come back as tool output.

Paid-only, key-scoped, revocable

One sgpt_… MCP key (workspace Settings → MCP) in the harness config — it reaches every agent in the workspace. Free-plan keys get the upgrade message, never an answer. Chat API keys never work on MCP.

Setup

Claude Code

Run the one-liner (replace <app-host> with your deployed host and sgpt_… with an MCP key) — or add the JSON to .mcp.json by hand.

terminalbash
claude mcp add --transport http helpgpt https://<app-host>/api/mcp --header "Authorization: Bearer sgpt_…"
.mcp.jsonjson
{
  "mcpServers": {
    "helpgpt": {
      "type": "http",
      "url": "https://<app-host>/api/mcp",
      "headers": {
        "Authorization": "Bearer sgpt_…"
      }
    }
  }
}

Codex

export HELPGPT_API_KEY=sgpt_… first, then run the one-liner — or edit ~/.codex/config.toml by hand.

terminalbash
codex mcp add helpgpt --url https://<app-host>/api/mcp --bearer-token-env-var HELPGPT_API_KEY
config.tomltoml
[mcp_servers.helpgpt]
url = "https://<app-host>/api/mcp"
bearer_token_env_var = "HELPGPT_API_KEY"

Cursor (IDE + CLI)

export HELPGPT_API_KEY=sgpt_… first, then add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global) — both the IDE and the cursor-agent CLI read this file (the CLI has no mcp add), then restart.

mcp.jsonjson
{
  "mcpServers": {
    "helpgpt": {
      "url": "https://<app-host>/api/mcp",
      "headers": {
        "Authorization": "Bearer ${env:HELPGPT_API_KEY}"
      }
    }
  }
}

OpenClaw

Same endpoint, same sgpt_… MCP key — add to ~/.openclaw/openclaw.json (or Settings → MCP → Add server → Streamable HTTP), replacing <app-host> with your deployed host:

openclaw.jsonjson
{
  "mcpServers": {
    "helpgpt": {
      "transport": "streamable-http",
      "url": "https://<app-host>/api/mcp",
      "headers": {
        "Authorization": "Bearer sgpt_…"
      }
    }
  }
}

No MCP client? Plain HTTPS

Any agent that can run curl can ask your docs. Use a chat-purpose key from the agent's API Keys page (MCP keys don't work on this route):

terminalbash
curl -N -X POST "https://<app-host>/api/chat/<agent-id>?channel=api" \
  -H "Authorization: Bearer sgpt_…" \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"id":"msg-1","role":"user","parts":[{"type":"text","text":"How do I reset my password?"}]}]}'

Usage

Once connected, call the tools from your harness. Replace <agent-id> with an id from helpgpt_list_agents.

1 · Find your agent

Call helpgpt_list_agents (no arguments) and copy the id of the agent you want to ask.

2 · Ask your docs

Call helpgpt_ask with your agentId and question — for example { agentId: "<agent-id>", question: "How do I reset my password?" }. The answer comes back with cited sources, or a clean refusal when the docs don't cover it.

3 · Use the free tools

Call any text-based tool directly — for example sitemap_validate with pasted XML, or convert_markdown with converter: "csv" and pasted content. Results return inline as tool output.

4 · Watch the cap

Asks and PDF AI tools count against your workspace message cap like widget turns; pure-local tools count nothing. A capped workspace gets a clear message instead of an answer.

Tools (12)

  • helpgpt_ask

    Grounded answer with citations or clean refusal.

  • helpgpt_list_agents

    Workspace agents (id + name) to ask.

  • convert_markdown

    Pasted text or page URL → Markdown.

  • sitemap_generate

    URL list → sitemap.xml.

  • sitemap_validate

    sitemap.xml → validation findings.

  • sitemap_extract_urls

    sitemap.xml → URL list.

  • page_links

    HTML + URL → internal/external links.

  • transcript_analyze

    Chat transcript → support metrics.

  • roi_calculate

    Ticket numbers → savings and ROI.

  • signature_generate

    Contact details → HTML signature.

  • pdf_ask

    Pasted PDF text + question → answer.

  • pdf_summarize

    Pasted PDF text → summary.

FAQ

Which plans include MCP access?

Starter, Growth, and Scale. Free-plan keys are refused with an upgrade message — the website's free tools stay free on the web for everyone.

Where do I get an API key?

In your workspace: go to Settings → MCP and create an MCP key. It is shown in plaintext once — store it in an env var, never in git. Chat API keys (per-agent page) do not work on MCP.

Does MCP usage count against my message cap?

Asks and PDF AI tools count like widget turns. Pure-local tools (converters, sitemap, analysis, ROI, signatures) count nothing.

Does OpenClaw work?

Yes — OpenClaw connects over the same Streamable HTTP endpoint with the same sgpt_… MCP key. See Setup — OpenClaw above. Paid workspaces only, same as every other MCP client.

My agent can't do MCP?

Use plain HTTPS: POST your question to /api/chat/<agent-id>?channel=api with a chat-purpose key (per-agent API Keys page) and the AI-SDK message shape. Same cited answers, same caps, no MCP client needed.

What stays out of MCP?

Anything needing a file upload, producing a downloadable file, or interactive — the PDF suite, OCR, audio, invoices, sharing/whiteboard, the sitemap finder. Source and agent management stay in the dashboard.

Already a customer? Open MCP in your workspace for the setup with your real host filled in.

Start answering from your own content

Start free