---
title: "MCP prompts"
description: "Reference for Alpacon's four MCP prompts, workflow guides that teach an agent to scope Work Sessions, execute guarded actions, handle incidents, and audit activity."
---

> Full documentation index: https://docs.alpacax.com/llms.txt

# MCP prompts

MCP prompts are reusable workflow guides that an MCP client can invoke by name, distinct from [tools](/reference/mcp/tools/) (which act) and [resources](/reference/mcp/resources/) (which expose data). Each Alpacon prompt returns static guidance text—arguments only scope the context inside it—that walks an AI agent through the discipline required to operate Alpacon safely: every infrastructure action happens inside an approved Work Session, approval always routes to a human because an agent has no MFA presence, and denials are structured so the agent self-corrects instead of retrying blindly. The four prompts mirror Alpacon's access model, ACCESS→EXECUTION→AUDIT.

  Prompts don't call the Alpacon API themselves—they return instructions that the agent then carries out using the corresponding tools. See [Getting started](/integrate/mcp/) for how your MCP client surfaces prompts.

## Summary

| Prompt | Stage | Arguments |
|--------|-------|-----------|
| `work_session_workflow` | ACCESS | `intent` (required), `servers` (optional) |
| `guarded_execution` | EXECUTION | `work_session_id` (required) |
| `incident_response` | Scenario | `server_id` (optional), `workspace` (optional) |
| `security_audit` | AUDIT | `work_session_id` (optional), `server_id` (optional) |

## work_session_workflow

Invoke this prompt before any infrastructure action, to scope and open a Work Session. Given a goal (`intent`) and optionally a list of target server UUIDs (`servers`), it walks the agent through:

1. Declaring the goal as the session `description` and calling `work_session_create` with the required `workspace`, `scopes`, `servers`, and `expires_at`, scoped to the specific target servers rather than the whole workspace. It clarifies which scopes an agent can request directly (`command`, `webftp`, `tunnel`) versus `sudo`, and which are unavailable to agents entirely (`websh`, `editor`, which require MFA presence). An agent can request `sudo`, but a sudo invocation that no Work Session sudo policy covers is denied at once. A human can clear some of these denials, after which the agent reruns the command; the rest leave nothing to approve.
2. Handling the result by status: a `pending_approval` session means the agent must surface the request to a human and wait rather than retry—`explain_approval_decision` can relay why a human must act out-of-band. An `error` with a gate `code` means reading `next_action`, narrowing the scope, and retrying deliberately.
3. Once the session is `active` (confirmed via `work_session_get`), handing off to the `guarded_execution` prompt for the actual work.

Use it whenever a user's request implies touching infrastructure—running a command, transferring a file, or opening a tunnel—and no Work Session is open yet.

## guarded_execution

Invoke this prompt once a Work Session (`work_session_id`) is active, to run commands and file transfers inside it correctly. It walks the agent through:

1. Confirming the session is `active` via `work_session_get`, reading `data.status` rather than the top-level call-result `status`, before doing anything.
2. Passing `work_session_id` on every scoped call: `execute_command`, `execute_command_multi_server`, or `webftp_download_file` (`webftp_upload_file` is local-mode only and returns `remote_mode_unsupported` over remote/OAuth).
3. Expecting a risk verdict on every action, and that the agent cannot self-approve. A HIGH-risk command is held as `pending_approval` and runs by itself once a human approves it. A `sudo` inside a command does not wait: when no Work Session sudo policy covers it, it is denied at once and the command finishes with the denial in its result. The agent surfaces a `sudo_denial` to a human and reruns the command after they act, and treats a `sudo_hint` with no `sudo_denial` as a hard denial with nothing to wait for.
4. Reading the structured `code`/`next_action` on denial and self-correcting instead of re-running the same denied command.
5. Calling `work_session_close` once the work is done, which also triggers the AI security analysis over the session.

Use it for every command or file transfer that happens right after a Work Session becomes active.

## incident_response

Invoke this prompt when responding to an alert or reported incident, to make sure triage happens before any remediation. Given an optional `server_id` or `workspace`, it walks the agent through:

1. Read-only triage that needs no Work Session: active alerts, server overview and CPU/memory usage, and recent events, to identify the likely cause first.
2. If remediation requires running commands, opening a Work Session through the `work_session_workflow` prompt, scoped to the affected servers only with the `command` scope, instead of executing directly.
3. Running the remediation under `guarded_execution`, expecting risky steps (restarts, destructive commands) to be held for human approval and a denied `sudo` to be handled as `guarded_execution` describes.
4. Closing the session with `work_session_close` once the incident is resolved, to complete the audit trail.

Use it when a user reports a problem with a server or asks the agent to investigate an alert.

## security_audit

Invoke this prompt when a user asks an audit or forensic question about privileged activity, to pick the right lens instead of guessing which tool to call. Given an optional `work_session_id` or `server_id` as anchor, it maps five question types to the tools that answer them:

- **Session forensic** ("what happened in this session?"): `work_session_timeline`.
- **Event forensic, cross-session** ("every sudo call", "every transfer of /etc/*"): `list_server_logs`, `list_webftp_logs`, `search_events`.
- **Decision audit** ("who approved this, when was it revoked?"): `list_approval_requests`, `get_approval_request`.
- **Mutation audit** ("what changed in roles, tokens, or policy?"): `list_activity_logs`, `get_activity_log`.
- **AI-derived analysis** (attack patterns, MITRE ATT&CK mapping): `list_session_analyses`, `get_session_analysis_detail`.

It also directs the agent to link findings back to the Work Session that contains each event, since the session is the primary unit of the audit story.

Use it for any "what happened," "who approved," or "was this suspicious" question about workspace activity.

## Related

- [Work Session tools](/reference/mcp/tools/work-sessions/): the tools these prompts orchestrate, including `work_session_create`, `work_session_get`, and `work_session_close`
- [Audit log tools](/reference/mcp/tools/audit/): the tools behind the `security_audit` lenses
- [Getting started](/integrate/mcp/): connect your MCP client and see how it surfaces prompts and resources

Source: https://docs.alpacax.com/reference/mcp/prompts/
