Skip to main content
Runlayer MCP (the platform self-MCP) is a first-party MCP server that lets an assistant ask questions about Runlayer itself: usage analytics, audit logs, server inventory, policies, plugins, and more. It is the fastest way to answer “what happened?” and “how is adoption going?” without clicking around the UI. Runlayer MCP is not the same as Runlayer Plugin. Installing Runlayer MCP does not expose your organization’s connector MCPs — it only exposes the curated set of platform tools described below (audit logs, analytics, inventory, governance). Runlayer Plugin is the separate unified entrypoint for connector tools, and even there, connectors are exposed conditionally: only active hosted connectors the user can access through policy, never local, draft, disabled, or policy-blocked connectors.
All Runlayer MCP tool calls are governed by your Runlayer permissions and show up in Audit Logs.

When to use Runlayer MCP

Use Runlayer MCP when you want to:
  • Pull analytics datasets for reporting (ex: top tools/users/servers over a date range)
  • Query audit logs to investigate failures, policy denials, or suspicious activity
  • Generate a server inventory (what MCPs exist, who can access them, what tools they expose)
  • Inspect governance state (policies, access requests, security scanner settings)
  • Create or maintain plugins, skills, and agents (admin / owner gated, confirm-required)

What it can do (high level)

Runlayer MCP intentionally exposes a curated set of high-leverage tools (not every backend endpoint). Common capabilities:
  • Audit logs
    • Discover valid action and resource type filter values before listing or exporting audit logs
    • List recent audit events with filters (action type, resource type/id, actor, correlation id, request id, client name)
    • Fetch a single audit log entry with full details when you need deep inspection
    • Bulk export audit logs as JSON or CSV — no 90-day hot-window restriction, up to 10 000 rows per call. Non-admins are scoped to their own logs; admins see org-wide logs. Sensitive-server details are redacted unless the caller has the viewer role
  • Analytics exports (admin-only)
    • Export dashboard datasets as CSV (same underlying metrics as the Analytics UI)
    • Export per-server tool health + usage breakdowns as CSV
    • Read the agent usage feed: every agent’s creator, consumers, sharing, and token usage (get_agent_usage_report, list_agent_usage_consumers; needs View company metrics and Export usage data)
    • Read the Usage page’s Spend tab breakdowns with the same numbers the UI shows: spend per provider and model (get_spend_cost_breakdown), the All users list (get_spend_people_usage), and spend by department (get_spend_department_usage). Each needs View company metrics and covers at most 31 days per call
  • Servers
    • List servers you can access
    • Get server details + generate MCP client install links (Cursor / VS Code)
    • List tools/resources/prompts for a server (via Runlayer proxy)
    • Create, update, and delete servers (admin-only, confirm-required) — create_server and update_server support the same fields as the REST API, including OAuth configuration (manual and broker), audit log sensitivity, tool output offloading, MCP fingerprint, and deployment association
  • Governance (mostly admin-only)
    • List, read, create, update, and delete policies (full CRUD — mutations require explicit confirm=true)
    • List access requests
    • Read/update security scanner settings (mutations require explicit confirm)
  • Incidents
    • List incidents with filters (status, priority, source type, date range) and pagination
    • Get full incident details including grouped entries, available triage/remediation actions, and activity history
    • Triage incidents (resolve, snooze, dismiss) and dispatch remediation actions (block tool, disable connector, escalate scanner) — mutations require explicit confirm=true
  • Sessions
    • List recent sessions with filters (status, client, user, date range) and pagination
    • Get session details including events, tool calls, scan results, and AgentGuard findings
    • Stop a session so its tool calls are denied, and clear the stop (stop_session, clear_session_stop). Stopping also cancels the session’s running Runlayer Agents runs, whoever started them. Both need Manage security policies, a reason, and confirm=true, and both are recorded in the audit log. Only a session Runlayer has recorded can be stopped; an unknown session id is refused
  • Automation building blocks
    • Plugins: create/update, manage connectors (tool allowlists) + attach skills
    • Skills: create/update skill instruction files used by plugins
    • Agents: manage agents; edit connectors and skills incrementally (set_agent_connectors / add_agent_connectors / remove_agent_connectors, set_agent_skills / add_agent_skills / remove_agent_skills); create schedules/webhooks; configure Slack DM mode; enable/disable memory; enable/disable agents; delete agents (delete_agent, owner or agent manager only, confirm-required). create_agent requires ai_provider_id and accepts optional model and icon parameters; update_agent covers metadata and settings such as is_disabled, and accepts ai_provider_id to move an agent to a different AI provider (pass it alongside model when the new model lives on a different provider).
    • Agent models: list available models for agents and update an agent’s model
    • Agent icons: create_agent and update_agent accept a predefined icon, an emoji, or an uploaded icon. generate_agent_icon draws an AI icon suggestion, the same as Generate in the icon picker, and returns a key you pass to update_agent to apply it. Needs edit access to the agent
    • Agent ownership: transfer_agent_owner makes another person in the organization the owner of an agent, with the same checks and warnings as Transfer owner in the Share dialog. Needs permission to transfer ownership and confirm=true. Agents can’t call this tool
    • Version history: list_agent_versions lists an agent’s versions, newest first, with what changed and who made each change. get_agent_version returns the full configuration one version saved, with environment variable names but never their values. restore_agent_version saves an earlier version as a new version, so the restore can be undone. It never changes environment variables or secrets, sharing, schedules or webhooks, and it returns connectors or skills deleted since that version as warnings. Pass expected_current_version to refuse with conflict if someone saved a newer version in the meantime. All three need Use Agents; listing and reading need view access to the agent, and restoring needs edit access and confirm=true
    • Usage limits: get_agent returns the agent’s spend and run caps with current usage and reset times. update_agent sets or removes the daily and weekly cost budgets; only agent managers can change limits
    • Schedules: run_agent_schedule_now runs a saved schedule once, now, like the schedule card’s Run now. It uses the schedule’s prompt and model with your connections and memory, shows under the schedule, and doesn’t move the schedule’s next or last run. Needs run access to the agent and confirm=true
    • Webhooks: update_agent_webhook renames a webhook. create_agent_webhook and set_agent_webhook_signing turn request signing on, rotate the signing secret, or turn signing off; a new signing secret appears in the tool response only once. An agent can’t turn on signing for its own webhooks
    • Poll triggers: list_agent_trigger_poll_attempts lists a trigger’s recent polls, poll_agent_trigger_now runs Fetch now (confirm=true), and list_agent_trigger_target_options lists the live choices for a trigger’s target fields, such as Slack channels or GitHub repositories. Needs edit access to the agent; only the owner of the trigger’s connected account can run Fetch now, see what a poll returned, or change what the trigger watches and who receives its runs
    • Evals: delete_agent_eval deletes an eval, set_agent_eval_human_verdict records your verdict on an eval case, and suggest_agent_eval_runs picks an eval’s source runs from a plain-language description. set_agent_run_annotation, delete_agent_run_annotation and list_agent_run_annotations manage pass/fail labels on runs, and analyze_agent_run_failures groups failed runs by cause. All need edit access to the agent; the changes need confirm=true
    • GitHub access: read an agent’s GitHub access and the org’s GitHub App connections (get_agent_github_access), grant a connection to an agent with an optional owner/repo list (grant_agent_github_access), and revoke it (revoke_agent_github_access). All three need Manage org settings; the writes need confirm=true and are recorded in the audit log. Agents can’t call the grant or revoke tool. Connecting the GitHub App itself is only in Settings → Integrations, because it takes the App’s private key
    • Subagents: add, update, and delete subagents on an agent (confirm-required)
    • Artifacts: create standalone files from text or base64 content, find and read accessible artifacts, update or edit standalone artifacts you own, and delete artifacts you can manage. Mutations require confirm=true. See Create and update for the workflow.
If you are unsure what to do next, ask your assistant to “use Runlayer MCP and tell me what tools you have available” and then pick the smallest set of tools needed for the task.

Getting started

1

Make sure the Runlayer server exists (admin)

Go to Connectors, find Runlayer, and click Add connector if it is not already enabled. If it is already installed, you should see a connector named Runlayer in the connectors UI.
2

Connect it to your AI client

In the Runlayer UI, open the Runlayer connector and click Add to Client (or copy the proxy URL). Then enable/select the connector in your AI client.
3

Ask a reporting question

Start with one of the example prompts below. If you need a CSV export, ask for a CSV and a short summary (the assistant can parse the CSV output).
Building an agent that calls Runlayer MCP programmatically? See Agent Accounts for OAuth client credentials + token flows.

Example prompts (copy/paste)

Analytics (adoption + usage)

  • “Use Runlayer MCP to export top tools used in the last 30 days as CSV, then summarize the top 10 and call out failure rate.”
  • “Export tool calls hourly for the last 7 days (UTC) and describe the daily pattern.”
  • “Which MCP clients (Cursor/Claude/etc) generated the most MCP activity last month? Export CSV and rank them.”

Audit logs (investigations)

  • “Use Runlayer MCP to list recent tool_call_failure audit logs. Group by server name and show the most common error messages.”
  • “Find audit logs with correlation_id=<id> and summarize the sequence of events.”
  • “For the most recent policy denied events, show actor + server + tool and the violation reason.”
  • “Export the last 30 days of security violation audit logs as CSV and identify the top offending tools.”

Incidents + sessions

  • “Use Runlayer MCP to list open incidents from the last 7 days and summarize by source type.”
  • “Get the details of the highest-priority incident and show the available remediation actions.”
  • “List sessions for user@company.com from the last 24 hours and flag any with security violations.”

Inventory + debugging

  • “List servers I can access, then for the Slack server list its tools and give me the install link for Cursor.”
  • “I think a connector is missing tools. List the server tools via Runlayer MCP and compare to what the UI shows.”

Notes and gotchas

  • Admin-only vs non-admin: some reporting and governance tools require admin access. Non-admins may see a reduced view (for example, audit logs are limited to the caller unless you are an admin).
  • Safe-by-default mutations: tools that change state require an explicit confirm=true parameter. If your assistant asks to “confirm”, that is expected.
  • Large outputs: some audit log entries can be huge. Listing tools may omit large fields; fetch the specific record when you need the full payload.
  • Changing agents from an agent: when Runlayer MCP runs inside a Runlayer agent’s own run (for example in Slack or the playground), that run can change only its own agent; changes to other agents fail with agent_scope_denied. Attach Runlayer MCP to the agent as a connector to let it manage other agents with the permissions of the user it acts for. Your own harness on an Agent Account gets no agent-changing tools by default, even though everyone can use Runlayer MCP: an admin must connect its agent account to the Runlayer connector with the specific tools, for example update_agent. With an OBO token it can then change only the agents its delegating user can edit; any other agent-changing tool fails with agent_account_not_granted. A grant covers a whole tool, not individual changes: update_agent can also replace an agent’s connectors and skills through its connectors and skill_ids parameters (it cannot transfer ownership), so grant the narrower tools, such as add_agent_connectors or add_agent_skills, when those are all the harness needs.

Analytics

UI analytics dashboard

Audit Logs

Activity history + investigations

Incidents

Grouped security inbox

Sessions

Session monitoring + enforcement

Agent Accounts

Programmatic auth for agents

MCP Troubleshooting

Debug MCP connection + permission issues