Ask your agency anything, from your AI agent

Things to ask

• “Which retainers are going to blow this month?”
• “What did we promise Acme, and is any of it at risk?”
• “What's at risk across our projects this week?”
• “Log 90 minutes on AC-12: homepage copy revisions.”
• “Which client requests need review? Send the Acme logo one back, it's the wrong format.”
• “Create a task for the Acme launch checklist and assign it to Priya.”

Set it up in two minutes

1. In ROWANe, open Settings → Integrations → MCP and create a key. Pick the modules it may read. Leave writes off unless you want the agent to change things. The key is shown once, with the exact commands below already filled in.
2. Add the server to your client.

Claude Code

claude mcp add --transport http rowane https://rowane.app/api/mcp \
  --header "Authorization: Bearer rwn_mcp_…"

Claude Desktop, Cursor and other clients (mcp.json)

{
  "mcpServers": {
    "rowane": {
      "type": "http",
      "url": "https://rowane.app/api/mcp",
      "headers": {
        "Authorization": "Bearer rwn_mcp_…"
      }
    }
  }
}

The endpoint speaks Streamable HTTP over POST, using JSON-RPC 2.0.

What keeps it safe

Never wider than you

Every request is checked against your current permissions, your key's module list and your organization's MCP policy. If you lose access to a client, the agent loses it on its next request.

Read-only by default

Writes need the key, the organization policy and your own task permissions to all say yes. Money is never writable: budgets, rates and invoices change only in the app.

Everything is on the record

Every tool call and every change is written to your organization's audit log against the key that made it.

Owners stay in control

Owners can turn MCP off org-wide, limit the modules, restrict keys to IP addresses and revoke any key. Revocation takes effect on the next request.

Client text stays data

Text written by clients is marked as untrusted before your agent sees it, so instructions hidden in a request can't pass as yours.

Some things stay human

Accepting files a client uploaded, and anything touching money, happen in the app.

Tools

A key is only offered the tools its modules allow. The rest don't exist for it.

Work · needs Tasks

• my_work — What's on your plate, by urgency
• search_tasks / get_task — Find and read tasks
• list_task_statuses — The columns a task can move to
• find_members — Look up a teammate to assign work to

Sprints · needs Sprints

• list_series / list_sprints / get_sprint_status — Sprint progress

Clients & promises · needs Clients

• list_clients / get_client — Clients you're staffed on
• list_commitments — What you promised each client, and whether it's at risk
• list_client_requests — Requests raised by clients, and answers waiting for review

Projects & docs · needs Projects / Documents

• list_projects / get_project — Projects and their state
• list_documents / list_folders / get_document — The file library

Money · needs Expenses

• list_budgets / get_budget_burn — Retainer and budget burn, this period

Delivery risk · needs Intelligence

• get_delivery_risks — Open risk cases, including missed and at-risk commitments
• trigger_delivery_scan — Re-scan a project or client now

Writes (only on a key with writes turned on) · needs Tasks, plus Clients for triage

• create_task / update_task / set_task_status — Change work
• add_task_comment — Comment (internal unless you ask to share it)
• log_work — Log your own time on a task
• triage_client_request — Resolve, decline, send back or cancel a client request