Skip to content

Platform tools#

Everything nexusctl and the REST API can do, the MCP endpoint exposes as tools. Connect a coding agent to the platform and administer it in natural language: "create a staging project", "which sandboxes went into error today", "cap this agent at fifty dollars a month".

The agent acts as you — your identity, your role, your policies, your rows in the audit trail.

Tool domains#

Domain Tools
Organizations list_orgs, get_org, create_org, update_org, delete_org, remove_org_member
Projects list_projects, get_project, create_project, update_project, delete_project, get_project_public_key, rotate_project_keys
Project members list_project_members, set_project_member, remove_project_member
Invitations invite_org_member, list_my_invitations, accept_invitation, decline_invitation
API tokens list_api_tokens, create_api_token, revoke_api_token
Sandboxes list_sandboxes, get_sandbox, create_sandbox, update_sandbox, start_sandbox, stop_sandbox, restart_sandbox, delete_sandbox, list_sandbox_revisions, rollback_sandbox, get_sandbox_effective_policy
Policies list_policies, list_org_policies, get_policy, create_policy, create_org_policy, update_policy, delete_policy
Policy bindings list_policy_bindings, list_org_policy_bindings, get_policy_binding, create_policy_binding, create_org_policy_binding, update_policy_binding, delete_policy_binding, list_policy_binding_drift
Credentials list_credentials, get_credential, create_credential, update_credential, delete_credential
Clusters list_clusters, get_cluster, create_cluster, update_cluster, delete_cluster, rotate_cluster_tunnel_token
AWS connections list_aws_connections, get_aws_connection, create_aws_connection, update_aws_connection, delete_aws_connection
Connectors list_mcp_servers, get_mcp_server, create_mcp_server, update_mcp_server, delete_mcp_server, sync_mcp_server_tools, create_http_connector, list_connector_entries, set_connector_entry_visibility, list_mcp_server_credentials, create_mcp_server_credential, delete_mcp_server_credential, list_oauth_applications
Inference keys list_inference_keys, set_inference_key, remove_inference_key, list_inference_providers
Spending limits list_spending_limits, set_spending_limit, remove_spending_limit, get_spending_limit_status
Usage costs get_usage_cost_summary, get_usage_cost_timeseries
Audit trail query_audit_trail, get_audit_stats, get_audit_timeseries

The inference-key tools manage the provider keys an organization sets for itself in place of the install's own; see Inference providers. Keys are write-only: list_inference_keys shows only their last four characters. Listing, setting, and removing keys requires a human org admin signed in over OIDC, never an API token or a sandbox, so an agent cannot point its own inference at a key it chose.

Call tools/list for the authoritative set and full parameter schemas on your install — that is what a connected agent reads, and it stays correct as the platform evolves.

What each identity sees#

The list above is what an interactively signed-in user gets. Other identities get less:

  • API tokens see the operational tools — sandboxes, policies, policy bindings, credentials, clusters, AWS connections, audit queries, usage costs, and spending-limit status — and can read projects, project members, connectors, OAuth applications, and the organization's inference providers (list_inference_providers). They do not see organization or project lifecycle, project member changes, invitations, API token management, connector changes, or spending-limit changes.
  • Sandboxes see only get_spending_limit_status, get_usage_cost_summary, and get_usage_cost_timeseries. An agent can ask what it has spent. It cannot administer the platform running it.

This is enforced when the tool list is built, not at call time, so restricted tools are not merely refused — they are invisible.

Give the agent the mental model#

Tool schemas describe parameters. They do not describe sequencing, or which of two valid orders is the one that works.

The Lens Agents admin skill carries that: the platform mental model, correct ordering, and the gotchas. Install it once and your coding agent applies it to whatever administration task you hand it. See Skills.

Guardrails still apply#

Driving the platform through an agent does not widen what you can do:

  • RBAC is unchanged. The agent cannot reach a project you cannot reach.
  • Policies still resolve the same way. Org policy ceilings clip project bindings whether a person or an agent created them.
  • Every call is audited with your identity attached, so an agent's administrative actions are reviewable line by line.

Use the admin UI to validate and observe what the agent did.