Guides

MCP server

Drive your Rykon AI company from the AI tools you already use. One URL, signed in with your Rykon account, inside the rules your company already has.

Claude Code

  1. Add the server

    Terminal
    claude mcp add --transport http rykon-ai https://mcp.rykon.ai/mcp
  2. Sign in

    In a Claude Code session, run /mcp, choose rykon-ai and Authenticate. Your browser opens your Rykon account; sign in as usual. Then choose which company the connection is for and what it may do (below), and Allow.

  3. Ask

    For example: “What is waiting for me in Rykon AI?” or “Give the Marketing team a task to draft three subject lines for the launch, then tell me when it is done.”

Without a browser (a CI job, a script), use a Rykon AI API key as the bearer instead. Keep the key in an environment variable and name it in the project's .mcp.json; Claude Code fills it in when it starts, so the key is never written into the file or onto a command line.

.mcp.json
{
  "mcpServers": {
    "rykon-ai": {
      "type": "http",
      "url": "https://mcp.rykon.ai/mcp",
      "headers": { "Authorization": "Bearer ${RYKON_AI_API_KEY}" }
    }
  }
}

Cursor

Add the server to ~/.cursor/mcp.json (or the project's .cursor/mcp.json), then open Cursor Settings, MCP, and choose Connect beside rykon-ai to sign in.

mcp.json
{
  "mcpServers": {
    "rykon-ai": { "url": "https://mcp.rykon.ai/mcp" }
  }
}

Claude Desktop

When the hosted service is open: Settings, Connectors, Add custom connector, and paste https://mcp.rykon.ai/mcp. Until then, connect from your own machine through the mcp-remote bridge in claude_desktop_config.json (Settings, Developer, Edit config), restart Claude Desktop, and sign in when the browser opens.

claude_desktop_config.json
{
  "mcpServers": {
    "rykon-ai": {
      "command": "npx",
      "args": ["-y", "mcp-remote@0.14.3", "https://mcp.rykon.ai/mcp"]
    }
  }
}

Signing in and what you allow

The server is its own OAuth authorization server and signs you in with your Rykon account, the same sign-in as the console. Your app registers itself (or presents a client id document), uses PKCE, and gets a token for this server only. On the consent page you choose:

  • Which company. A connection works on one company. To use another, connect again and choose it.
  • Read: the company, teams, agents, goals, tasks, runs and transcripts, documents, and what waits for you.
  • Read and start work: also add tasks for a team and start runs. Offered to owners, admins and members; viewers can read only.

The connection acts as you, never more than your role allows, and every tool call is on the account's audit log with the app that made it. Access tokens last an hour and refresh on their own; a refresh token is used once, and a stolen copy used again ends the connection. End a connection any time under API keys, Connected apps, in the console; removing someone from the account ends theirs at once.

It never answers for you

No tool answers a decision or approves an action, and a connection cannot be given that permission. What waits for you comes back with its console link, where you answer it. An agent approving its own work is exactly what the approvals exist to prevent.

It can stop spending, never start it

An owner or admin can pause the company or a team, or switch it to Quiet, from a connected app. Resuming, going back to Active and changing a budget are done by a person in the console: a connection cannot be given that permission, so an app, or an agent steering one, can never turn the spending back on.

Tools

ToolNeedsWhat it does
list_companiesreadThe companies you belong to, your role, and which one this connection works on.
get_companyreadMission, autonomy ceiling, teams, goals, what waits for you, what it did recently.
list_teamsreadEach team with its manager, people (id, role, level), goals, work and spend.
get_teamreadA team: the manager’s latest plan, its tasks, runs, decisions and documents.
list_agentsreadEvery agent: title, team, role, level, model, tools.
list_goalsreadCompany and team goals with their measure, target and progress.
list_departmentsreadThe structure as a tree: departments, teams and sub-teams with leads, people, budgets and open work.
get_goal_treereadObjectives, the department goals under them and the key results under those, with owners and progress.
propose_redesignstart workAsk the designer for a redesign draft; a person reviews and applies it in the console.
get_designreadA design or redesign draft: what it adds, changes and archives and why, what is not allowed, its cost.
list_tasksreadRecent tasks by state, team and role, with their results.
get_taskreadA task with its runs, the decisions it raised and its documents.
add_taskstart workGive a team work; its manager schedules it and the named role does it.
start_runstart workStart one run for an agent, or a bare run on a model you name.
list_runsreadRuns, newest first, paged with a cursor.
get_runreadState, which provider and model answered, tokens, cost, result.
get_transcriptreadWhat happened in a run, in order; page with the cursor while it is live.
list_tool_callsreadEvery tool call of a run, its risk class, decision and outcome.
list_documentsreadDocuments the agents wrote, newest first, paged with a cursor.
get_documentreadOne document in full, with where it came from.
list_inboxreadDecisions and approvals waiting for you, each with its console link.
get_spendreadThis month’s AI spend against the budget, by team and by the model that answered; Active, Quiet or Paused.
pause_companystart workStop the company spending: no plans, no new runs; running work finishes or stops. Owners and admins. Resuming is in the console.
pause_teamstart workPause one team the same way. Owners and admins. Resuming is in the console.
quiet_companystart workLower spend without stopping: plans at most once a day, no scout. Back to Active is in the console.
kill_switch_statusreadWhether anything is stopped, by whom and why.

Runs and documents are also resources, rykonai://runs/{run_id} and rykonai://documents/{document_id}. Lists page with a cursor the previous answer returns. A refusal or a missing row comes back as a tool result in plain words (“this agent is archived”, “Rykon AI is stopped for this account”), never a stack trace.

Protocol

Streamable HTTP at https://mcp.rykon.ai/mcp, stateless: protocol revision 2026-07-28 (per request metadata and server/discover) and the earlier revisions 2025-11-25, 2025-06-18 and 2025-03-26 (initialize). Discovery is the protected resource metadata at /.well-known/oauth-protected-resource/mcp and the authorization server metadata at /.well-known/oauth-authorization-server; scopes are read and work. A token without work asked to start work gets 403 insufficient_scope, so your client can ask you again.