Skip to content
Documentation

MCP server (connect your AI assistant)

CHAMPREP hosts a remote Model Context Protocol (MCP) server. Connect it to any MCP-capable AI client — Claude (claude.ai and desktop), ChatGPT, Claude Code, Codex, Cursor, VS Code, Windsurf, and others — and your assistant can act on your CHAMPREP account: list and book calendar events, read and draft mail, manage work boards, search your drive, create invoices, notes, documents, and spreadsheets, and much more, all by conversation.

There is nothing to install for modern clients. The server is hosted at:

https://ai-api.champrep.com/mcp/v1

The assistant’s reasoning runs in your AI client with your client’s model — CHAMPREP MCP calls consume none of your CHAMPREP AI credits. Every tool call is executed with your own identity and is bounded by your plan, your organization’s policies, and (for API keys) the scopes you chose.

There are two ways to connect. Both are fully supported — pick by situation:

Browser sign-in (OAuth)API key
Best forPeople using an AI app interactivelyScripts, CI, servers, headless tools
SetupPaste the URL; approve in your browserCreate a key once; paste it into config
Secrets to handleNone — no key ever touches the client configYou store the cpk_… key in the client’s config
Access breadthYour full account (like being signed in)Only the scopes you grant the key (e.g. just calendar:read)
Works inclaude.ai, Claude Desktop, ChatGPT, Claude Code, Codex, Cursor, VS Code, WindsurfEverything, including clients with no browser
ExpirySessions refresh automatically; revoke anytimeOptional expiry date; revoke anytime
Revoke fromDisconnect in the clientAccount → API Keys on champrep.com

Rule of thumb: if a human is in front of the client, use browser sign-in. If a machine is, use an API key with the narrowest scopes that do the job.

Section titled “Option A: Sign in with your browser (recommended)”

Add the server URL with no key or headers. Your client discovers the CHAMPREP sign-in flow automatically (OAuth 2.1 with dynamic registration), opens your browser to a CHAMPREP consent page, and connects after you approve. If you’re already signed in to CHAMPREP in that browser you go straight to the approval screen; if you’re signed in to several CHAMPREP accounts, you pick which one to connect.

  1. Settings → Connectors → Add custom connector.
  2. Paste https://ai-api.champrep.com/mcp/v1 and save.
  3. Your browser opens the CHAMPREP consent page — review and click Allow.
  4. In any chat, enable the CHAMPREP connector from the tools menu and just ask: “Check my CHAMPREP mail for unread messages.”

Connectors require a ChatGPT plan with developer mode (Settings → Connectors → Advanced → Developer mode).

  1. Settings → Connectors → Create (or Add connector).
  2. Name it CHAMPREP, set the MCP server URL to https://ai-api.champrep.com/mcp/v1, choose OAuth authentication, and save.
  3. Approve the CHAMPREP consent page when the browser opens.
Terminal window
claude mcp add --transport http champrep https://ai-api.champrep.com/mcp/v1

Then run /mcp, select champrep, and choose Authenticate — your browser opens the consent page.

Settings → MCP → Add new MCP server, or add to ~/.cursor/mcp.json:

{
"mcpServers": {
"champrep": { "url": "https://ai-api.champrep.com/mcp/v1" }
}
}

When Cursor reports the server needs authentication, follow the sign-in prompt.

Add an HTTP MCP server with URL https://ai-api.champrep.com/mcp/v1 in the client’s MCP settings. Both support the OAuth flow: the first connection opens the CHAMPREP consent page.

  1. Sign in at champrep.com and open Account → API Keys.
  2. Create a key and select its scopes — these decide which tools the assistant can use. calendar:read exposes read-only calendar tools; calendar:write also exposes tools that create or change events. Grant only the services you want reachable.
  3. Copy the key (cpk_live_…). It is shown once.

Claude Code:

Terminal window
claude mcp add --transport http champrep https://ai-api.champrep.com/mcp/v1 \
--header "Authorization: Bearer cpk_live_YOUR_KEY"

Cursor / VS Code / Windsurf (mcp.json-style config):

{
"mcpServers": {
"champrep": {
"url": "https://ai-api.champrep.com/mcp/v1",
"headers": { "Authorization": "Bearer cpk_live_YOUR_KEY" }
}
}
}

Codex CLI (~/.codex/config.toml) — Codex talks to stdio servers, so use the bridge package:

[mcp_servers.champrep]
command = "npx"
args = ["-y", "@champrep/mcp"]
env = { CHAMPREP_API_KEY = "cpk_live_YOUR_KEY" }

Claude Desktop (stdio config) and other stdio-only clients — the @champrep/mcp bridge forwards a local stdio server to the hosted endpoint:

{
"mcpServers": {
"champrep": {
"command": "npx",
"args": ["-y", "@champrep/mcp"],
"env": { "CHAMPREP_API_KEY": "cpk_live_YOUR_KEY" }
}
}
}

If CHAMPREP_API_KEY is missing, the bridge opens your browser at the API Keys page and prints instructions instead of failing silently.

Raw HTTP (for your own integrations — plain JSON-RPC 2.0). Pass the key from an environment variable rather than inline, so it stays out of your shell history:

Terminal window
export CHAMPREP_API_KEY="cpk_live_..." # from Account -> API Keys
curl https://ai-api.champrep.com/mcp/v1 \
-H "Authorization: Bearer $CHAMPREP_API_KEY" \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Your MCP client controls whether, when, and how it asks you to confirm a tool call. The MCP specification’s user interaction guidance recommends human oversight and confirmation prompts, but does not mandate a particular interaction model. Prompting depends on the client and its settings, including any remembered permissions or automatic-approval rules.

CHAMPREP supplies tool annotations to help clients assess each operation:

  • Read-only tools (list events, search mail, read notes …) have readOnlyHint: true.
  • Tools that change things (send mail, create events, upload files …) have readOnlyHint: false.
  • Destructive tools have destructiveHint: true and remain hidden unless destructive access is enabled (next section).

Tool annotations describe behavior. They do not enforce a confirmation dialog or guarantee a prompt before each call, even for destructive tools. Review your client’s approval settings before allowing changes.

For ordinary MCP actions, CHAMPREP relies on the client’s consent handling rather than requiring a second CHAMPREP approval card. The server does not independently verify that the client displayed a prompt.

CHAMPREP still enforces connection restrictions, applicable API-key scopes, service authorization, organization policies, and rate limits. Actions routed through CHAMPREP’s action executor also undergo safety checks and are recorded in your account’s Activity log, with Undo where supported. Client consent cannot override a server-side denial or an organization-required approval; actions requiring CHAMPREP-specific approval may be refused over MCP.

By default the server hides tools that permanently delete or clear data — deleting drive folders, purging notes, removing records, and similar. If you have real cleanup to do and want your assistant’s help, the connection must permit destructive actions and the client must request them by adding ?destructive=true to the server URL:

https://ai-api.champrep.com/mcp/v1?destructive=true

Use that URL anywhere a URL is accepted (claude.ai connector, Claude Code, Cursor, ChatGPT). For the stdio bridge, set an environment variable instead:

"env": {
"CHAMPREP_API_KEY": "cpk_live_YOUR_KEY",
"CHAMPREP_ENABLE_DESTRUCTIVE": "true"
}

The URL flag or bridge setting alone does not override the connection’s permissions. A read-only connection, a disallowed service, or disabled destructive access still prevents the tool from being available.

You are responsible for what your assistant does with these tools. Deleted data may not be recoverable. Destructive tools still require the appropriate API-key scopes, such as a service’s :write scope or its separate delete scope where applicable. They still carry destructiveHint, but whether you see a confirmation prompt depends on your client and its configuration. The server-side checks and action audit described above still apply. Some destructive actions support Undo from your account’s Activity page — but treat every deletion as final.

Regardless of this flag, some things are never available over MCP: administrative operations, code-execution tools, and account deletion. The only exceptions to the code-execution rule are starting a QA preview build or a QA cloud IDE session: both run in QA’s isolated sandbox, after your approval (see MCP for Code, CI, Runner Farm and QA).

Tools cover the whole platform, always as you: Calendar (events, bookings, availability), Mail (read, search, draft, send, labels, rules), Drive (search, files, folders, sharing), Notes, Meetings (schedule, summaries, recordings), Work (boards, tickets), Docs (documents, comments, suggested edits, versions, export), Sheets (workbooks, ranges, formulas, formatting, validation, conditional formatting, protection, comments, versions, export), Invoices, Forms, Contacts, Charts, CHAMPREP Code, Runner Farm, QA, and Learn. Exactly which tools appear depends on your plan, your organization’s policies, your per-service AI permissions, and (for API keys) the key’s scopes.

For CHAMPREP Code (projects, clone plans, pipelines, job logs, runners), Runner Farm and QA, see MCP for Code, CI, Runner Farm and QA: what the assistant can do, example prompts, what needs your approval and what stays in the web apps.

Try these first:

  • “What’s on my CHAMPREP calendar this week?”
  • “Do I have unread mail? Summarize anything important.”
  • “Find the PDF contracts in my drive from this month.”
  • “Create a work ticket titled ‘Renew SSL certs’ on my Ops board.”
  • “Create a Docs document called ‘Q4 plan’ with sections for goals, risks, and owners.”
  • “In my ‘Monthly budget’ workbook, highlight every expense over 500.”

Browser Relay Bridge (Chrome extension) and Eve AI

Section titled “Browser Relay Bridge (Chrome extension) and Eve AI”

In addition to cloud service tools, the CHAMPREP for Chrome extension can connect your own browser to Eve:

  • Turn on Let Eve use this browser in the extension settings.
  • Eve chats on the web, in the desktop app, in Eve AI for VS Code, and in the terminal agent (eve / champrep ai) can then read the active tab, list tabs, and capture screenshots, and, after you approve each action, navigate, click, or fill in a field.
  • Eve never acts on mail, banking, payment, or sign-in sites and never types into password, one-time-code, or card fields.
  • API keys — view, scope, expire, and revoke at Account → API Keys. Revocation takes effect immediately.
  • Browser sign-ins — disconnect the connector inside the AI client. You can also sign out of CHAMPREP in that browser; new tokens can then no longer be issued for it.
  • Every agentic action appears in your account’s Activity log with what ran, when, from where — and Undo where supported.
  • 120 MCP requests per minute per identity. Heavier automation should use the developer API with service-scoped keys.
  • Tool calls hit the same per-plan quotas as the rest of your account — an assistant can’t exceed what you can do yourself.
SymptomLikely cause and fix
Client says authentication required and nothing opensYour client may not support MCP OAuth — use an API key (Option B).
Invalid API keyThe key was revoked, expired, or mistyped. Create a fresh one at Account → API Keys.
Unknown or unavailable toolThe tool is outside your key’s scopes, disabled by your AI or connection permissions, or destructive without both connection permission and the ?destructive=true opt-in.
A tool runs without a confirmation promptPrompting is controlled by your MCP client and its approval settings. Tool annotations, including destructiveHint, do not guarantee a prompt. Review remembered permissions and automatic-approval rules in that client.
A write tool is refused with a limit messageCHAMPREP’s server-side action gates (daily caps, abuse protection) declined it — the message says why. Try later or reduce volume.
Consent page says “Session expired”The approval sat too long. Re-run the connect flow from the client.
Works in browser test but not from a server/CIUse an API key — browser sign-in mints tokens only interactively.
429 responsesYou hit the per-minute ceiling; back off and retry.

Endpoint: https://ai-api.champrep.com/mcp/v1.