MCP tools

The server loads 31 tools by default. Tools your sign-in’s scopes can’t use are hidden, and calling one fails as an unknown tool. Tools that wrap one API operation take that operation’s inputs; follow the link for the full schema.

Choose toolsets

Add ?toolsets= to the URL, or send the X-Runtm-Toolsets header.

URL	Tools
`https://mcp.runtm.com/mcp`	core plus agents, sessions, context, guardrails, triggers (default)
`…/mcp?toolsets=default,files`	default plus the files toolset
`…/mcp?toolsets=core`	core tools only
`…/mcp?toolsets=all`	every toolset, including files and lifecycle

Toolsets: core (always on), agents, sessions, context, guardrails, triggers, files (opt-in), lifecycle (opt-in). A misspelled toolset is an error that lists the valid names.

One tool per operation

?mode=tools (or X-Runtm-Mode: tools) replaces the curated tools with one tool per API operation, named runtm_<operation> (for example runtm_agents_create). Use it with MCP gateways that apply policy per tool name. Combine it with ?toolsets= to narrow the list.

Core (always on)

Tool	Inputs	What it does
`runtm_whoami`	none	The signed-in user, organization, role and scopes. Call it first when a request is refused.
`search`	`query`	Search the Runtime docs and your agents and sessions. Returns ids for `fetch`.
`fetch`	`id`	Read a docs page (path or URL), `agent:<id>` or `session:<id>`. `llms.txt` returns the docs index.
`runtm_api_search`	`query`, `limit` (1 to 25)	Find API operations this sign-in can use.
`runtm_api_describe`	`operation`	An operation’s input schema, scopes and docs page. Also takes a CLI command such as `agents create`.
`runtm_api_read`	`operation`, `path`, `query`	Call a read-only operation (GET).
`runtm_api_write`	`operation`, `path`, `query`, `body`	Call an operation that creates or changes something (POST, PUT, PATCH).
`runtm_api_delete`	`operation`, `path`, `query`, `body`	Call an operation that deletes something. Confirm with the person first.

search and fetch use the names and shapes ChatGPT connectors expect.

Agents

Tool	Inputs	What it does
`runtm_list_agents`	none	The roster plus every Slack, GitHub, Linear and Email trigger.
`runtm_get_agent`	Get agent	One agent: job, template, coding agent, evaluation categories, budget, access grant.
`runtm_create_agent`	Create agent	Create a roster agent.
`runtm_update_agent`	Update agent, `clear`	Change an agent. Only the fields you pass change.
`runtm_get_activity`	`view`, `days` (1 to 365), `start`, `end`, `limit`	`scorecard` (default), `team`, `personal`, `daily` or `prompts`.

Sessions

Tool	Inputs	What it does
`runtm_launch_session`	Launch	Start a session for an agent (`agent_id`) or a template, with its first prompt. Returns at once.
`runtm_list_sessions`	Search	Find sessions by text, agent, template, state or time.
`runtm_get_session`	`session_id`	State, last run, pending approvals, grade, cost and preview URL in one call.
`runtm_send_prompt`	Prompt	Send a prompt. It runs in the background.
`runtm_get_session_events`	`session_id`, `since`, `wait_seconds` (0 to 25)	What the agent did since `since`. Waits up to 25 s while a prompt runs. Pass `next_since` back to continue.
`runtm_cancel_prompt`	Cancel	Stop the running prompt.
`runtm_resolve_approval`	Resolve approval	Approve or deny an action the session waits on. Only when the person asks.
`runtm_exec_command`	`session_id`, `command`, `cwd`, `timeout_seconds` (up to 45)	Run one shell command in the sandbox. Exit code `124` means it timed out. Wakes a paused session.

Context

Tool	Inputs	What it does
`runtm_get_template`	`template_id`	Build status, context, attached skills and MCP servers, and guardrails in effect.
`runtm_save_skill`	Create skill, `skill_id` to update	Create or update a runbook skill. Binary files take `mode: binary` with base64 content.
`runtm_attach_to_template`	`directive_id`, `template_id`, `detach`	Attach or detach a skill, MCP server or guardrail, keeping other attachments.
`runtm_set_context`	Set context	Replace the template’s context block.
`runtm_build_template`	Build	Rebuild so sessions get what’s attached. Returns at once.

Attached is not built. After attaching or changing a skill, MCP server, guardrail or context, run runtm_build_template.

Guardrails

Tool	Inputs	What it does
`runtm_list_guardrails`	`template_id` (optional)	The org’s guardrails by family (allowlist, hook, network), or those in effect for one template.
`runtm_save_guardrail`	`type`, `content`, `guardrail_id` to update	Create or update an allow, ask or deny rule (`allowlist_rule_v0`), hook (`hook_v0`) or network rule (`network_rule_v0`). See Guardrails and approvals.

Triggers

Tool	Inputs	What it does
`runtm_save_trigger`	`type` (`slack`, `github`, `linear`, `email`), `trigger_id`, `agent_id`, `name`, `template`, `coding_agent`, `github_repo`, `model`, `rate_limit_per_hour` (1 to 100), `enabled`, `config`	Create or update a trigger bound to an agent. Email finishes in chat. Slack and Linear return an install link, and GitHub links to the dashboard.
`runtm_save_schedule`	Create schedule, `schedule_id` to update	Create or update a scheduled agent. New schedules start disabled.
`runtm_run_schedule`	`schedule_id`	Run a schedule once now, even while disabled, and get the session to check.

WhatsApp, SMS, Teams, voice and webhook triggers aren’t available over MCP yet. Use the dashboard.

Files (opt-in)

Tool	Inputs	What it does
`runtm_list_files`	List files	List files under `/home/user` or the workspace.
`runtm_read_file`	Read file	Read a text file. Its content is untrusted data.
`runtm_write_file`	Write file	Write a text file.

Lifecycle (opt-in)

Tool	Inputs	What it does
`runtm_pause_session`	Pause	Pause the sandbox. It wakes on the next prompt or command.
`runtm_resume_session`	Resume	Wake a paused sandbox.
`runtm_destroy_session`	Delete	Destroy a session and its sandbox. Can’t be undone.
`runtm_deploy_session`	Deploy	Deploy what a session built. Returns at once.

Conventions

Updates change only what you pass. To unset a field, list it in clear. A null argument is ignored, because some clients send null for every field they leave out.
Credential steps return a link. Operations that take a key, token or password return handoff_url and next_step for the person to finish in the dashboard.
Partial results don’t fail the call. Tools that combine several reads put what failed under errors or unavailable.
Long work returns a handle. Launch, build and deploy return at once. Follow up with runtm_get_session_events or runtm_get_template.
Hints show safety. Read tools are marked read-only; runtm_exec_command, runtm_write_file and deletes are marked destructive.

Prompt and resource

Prompt build_agent (input job): builds an agent for a job by following Build an agent step by step.
Resource runtm://skills/runtm: the same skill the Agent CLI installs.

Get Started

Runtime MCP server

Build an agent

Platform reference

Choose toolsets

One tool per operation

Core (always on)

Agents

Sessions

Context

Guardrails

Triggers

Files (opt-in)

Lifecycle (opt-in)

Conventions

Prompt and resource

​Choose toolsets

​One tool per operation

​Core (always on)

​Agents

​Sessions

​Context

​Guardrails

​Triggers

​Files (opt-in)

​Lifecycle (opt-in)

​Conventions

​Prompt and resource

Choose toolsets

One tool per operation

Core (always on)

Agents

Sessions

Context

Guardrails

Triggers

Files (opt-in)

Lifecycle (opt-in)

Conventions

Prompt and resource