> ## Documentation Index > Fetch the complete documentation index at: https://docs.linkedcamp.com/llms.txt > Use this file to discover all available pages before exploring further. # AI agents > Agents that reply to prospects, draft content, and book meetings, plus the knowledge base they draw on. Requires the **Agents** permission. Tools in this group are hidden from `tools/list` and refused on call when your plan does not grant it. | Tool | Kind | | - | - | | [`agent_delete`](#agent_delete) | Write, Destructive | | [`agent_duplicate`](#agent_duplicate) | Write | | [`agent_get`](#agent_get) | Read-only | | [`agent_list`](#agent_list) | Read-only | | [`agent_logs`](#agent_logs) | Read-only | | [`agent_query`](#agent_query) | Read-only | | [`agent_stats`](#agent_stats) | Read-only | | [`agent_train`](#agent_train) | Write | | [`agent_upsert`](#agent_upsert) | Write, Destructive | | [`kb_attach`](#kb_attach) | Write | | [`kb_delete`](#kb_delete) | Write, Destructive | | [`kb_list`](#kb_list) | Read-only | | [`kb_upsert`](#kb_upsert) | Write, Destructive | ## agent\_delete Write Destructive Marked destructive in the protocol — your MCP client asks for confirmation before running it. Delete an agent. Campaigns and conversations pointing at it lose their automated replies, silently — prospects mid-conversation simply stop getting answers. Irreversible, and its knowledge attachments and training go with it. Read `agent_list` for which campaigns use it, show the user, and get an explicit yes first. To stop an agent replying without destroying it, use `agent_upsert` with `isActive: false`, or `conversation_update` with `assistant: false` for one thread. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. ## agent\_duplicate Write Copy an agent, optionally into a sub-account. Copies the configuration; the duplicate still needs `agent_train` before its knowledge works. Passing `userId` to place it in another user's account is accepted only from an account owner. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. Sub-account owner id. Omit to act on your own data. ## agent\_get Read-only One agent's full configuration: personality, intent, instructions, FAQs, follow-ups and booking action. Read this out to the user when they are choosing an existing agent. Configuration only: for what it actually said use `agent_logs`, for how it would answer a specific question `agent_query`, and for volumes `agent_stats`. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. ## agent\_list Read-only AI agents configured for this account. CHAT agents reply to prospects in the inbox, CONTENT agents draft LinkedIn posts, and the rest are specialised assistants. Start here when the user asks about automated replies, and show them `agent_get` so they can judge the fit before anything is attached. An agent writes replies; a persona (`persona_list`) decides who qualifies; a knowledge base (`kb_list`) is the material an agent draws on. Attaching one to a campaign is `campaign_update` — and it answers real prospects by itself, so never attach one unasked. Requires the "Agents" permission on the plan. ### Parameters Filter by type. Free-text search. Include conversation and booking counts. ## agent\_logs Read-only What an agent was asked and what it answered — the trace to read when its replies look wrong. What it actually said in real conversations. To try a hypothetical instead, without sending anything, use `agent_query`; to read the prospect-facing thread use `conversation_get`. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. 1-indexed page number. ## agent\_query Read-only Run a question through an agent against a real conversation, to see how it would answer. Returns the answer without sending anything — nothing reaches the prospect, which makes this the safe way to vet an agent before it goes active. For what it has already said use `agent_logs`. This is not a way to reply to anyone: sending an inbox message has no tool at all. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. Context for the retrieval. The question to put to the agent. ## agent\_stats Read-only Conversations handled, replies sent and appointments booked, across all agents. Totals across every agent, not one — there is no per-agent breakdown here. To judge a single agent's quality read `agent_logs`. Requires the "Agents" permission on the plan. *No parameters.* ## agent\_train Write Rebuild the agent's knowledge embeddings from its attached knowledge-base entries. Required after every `kb_upsert` or `kb_attach`, or the agent answers from stale material — this is the step most easily forgotten. It only re-indexes existing knowledge: it does not change the agent's instructions (`agent_upsert`) and does not teach it anything not attached via `kb_attach`. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. ## agent\_upsert Write Destructive Marked destructive in the protocol — your MCP client asks for confirmation before running it. Create an AI agent, or update one by passing `agentId`. A CHAT agent will reply to real prospects once it is active and attached to a campaign — set `isActive: false` while iterating on its instructions. The behaviour comes from `personality` (who it is), `intent` (what it is trying to achieve) and `instructions` (how it should write). This creates or edits the agent and nothing else. It does not attach it to a campaign (`campaign_update`), give it source material (`kb_upsert` then `kb_attach`), or build its embeddings (`agent_train`, which must run after any knowledge change). Test with `agent_query` before it touches a real conversation. There is no tool to delete an agent. Requires the "Agents" permission on the plan. ### Parameters Omit to create a new agent. Internal name. CHAT for inbox replies, CONTENT for post drafting. Company the agent represents; injected into its prompt. Primary model provider. Must be connected. Used when the primary provider errors. Language code, e.g. "en", "fr", "de", "it", "pt", "ja". Where the agent operates, e.g. \["linkedin","email"]. Who the agent is and how it carries itself. The outcome it is steering the conversation toward. Concrete writing rules: tone, length, what to avoid. Raw system prompt override. Usually leave unset. Subjects the agent is allowed to discuss. Canned answers the agent should use verbatim. A question leads ask. The answer the agent gives, verbatim. Follow-up messages sent when a lead goes quiet. Words the agent must never use. Field name is spelled this way in the schema. Minutes to wait before replying, so it reads as human. Cap on agent messages per conversation. Stand down as soon as a human replies in the thread. Situations in which the agent stops replying. Stop when the lead becomes hostile. Stop when the lead signs off. Stop when the lead declines. Enable calendar booking. Booking settings. Used when `action` is "APPOINTMENT". Which calendar to book into: a Calendly event, a GoHighLevel calendar id, or "CUSTOM" to share `customCalendarLink`. e.g. "calendly", "ghl". Booking page URL the agent shares. Required when `calendar` is "CUSTOM", which also forces `sendLinkOnly`. Send the booking link instead of negotiating a time. Stop replying to a lead once they have booked. Share this agent with these sub-accounts. Notify you when the agent replies. Whether the agent runs. Keep false while drafting. ## kb\_attach Write Include or exclude a knowledge entry for one agent. Run `agent_train` afterwards or the change has no effect on answers. Scoped to one agent: it neither creates the entry (`kb_upsert`) nor deletes it, and there is no tool to delete a knowledge entry. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. Knowledge-base entry id, from `kb_list`. true attaches, false detaches. ## kb\_delete Write Destructive Marked destructive in the protocol — your MCP client asks for confirmation before running it. Remove a knowledge entry and its embeddings. The source material is gone — a crawled page would have to be crawled again, pasted text retyped — so confirm with the user first; `kb_list` shows what is there. The agent keeps answering, just without this material, and it will still cite it until `agent_train` reruns. To stop one agent using an entry while keeping it, use `kb_attach` to exclude it instead of deleting. Requires the "Agents" permission on the plan. ### Parameters AI agent id, from `agent_list`. Knowledge-base entry id, from `kb_list`. ## kb\_list Read-only Source material agents can draw on: crawled URLs, uploaded files and pasted text. Account-wide. Which agent may use which entry is set per agent by `kb_attach`, so an entry listed here is not necessarily in use. Requires the "Agents" permission on the plan. *No parameters.* ## kb\_upsert Write Destructive Marked destructive in the protocol — your MCP client asks for confirmation before running it. Add source material for agents. Type URL crawls a page, TEXT stores pasted content. Three steps, and skipping either of the last two leaves the agent unchanged: add it here, attach it with `kb_attach`, then rebuild with `agent_train`. Storing material does not change how the agent writes — that is `agent_upsert` instructions. Requires the "Agents" permission on the plan. ### Parameters Omit to create a new entry. Filter by type. Short label. Required when type is URL. Required when type is TEXT. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.