MCP

Connect Siesta AI to any MCP-compatible AI client (Claude, Cursor, VS Code, Antigravity, and others). Once connected, the client can manage your agents, skills and workflows, send messages, and read conversations directly from your Siesta AI organization.

How to install

Add one of the connection definitions below to your AI client's MCP configuration. Replace [YOUR-API-KEY] and [YOUR-ORGANIZATION-KEY] with the values from your Siesta AI organization.

HTTP

Recommended. No external dependencies, though a few clients may not support it yet.

"siesta": {
  "type": "http",
  "url": "https://api.siesta.ai/mcp",
  "headers": {
    "X-Api-Key": "[YOUR-API-KEY]",
    "X-Org-Id": "[YOUR-ORGANIZATION-KEY]"
  }
}

OAuth

Recommended for interactive MCP clients that support OAuth. The client opens Siesta AI for sign-in and consent, so the MCP definition contains no API key headers.

"siesta": {
  "type": "http",
  "url": "https://api.siesta.ai/mcp"
}

npx (fallback)

Needs Node.js. Use only as a fallback when the client cannot connect over HTTP.

"siesta": {
  "command": "npx",
  "args": [
    "-y", "mcp-remote", "https://api.siesta.ai/mcp",
    "--header", "X-Api-Key:[YOUR-API-KEY]",
    "--header", "X-Org-Id:[YOUR-ORGANIZATION-KEY]"
  ]
}

Easiest way to install

Every AI provider (Claude, Antigravity, and others) stores MCP servers differently and expects these definitions in its own JSON config. The fastest path is to copy the JSON above (with your API key and organization ID from Siesta AI), paste it into the client's chat, and ask the AI to install it. Then follow any instructions it gives, for example restarting the client to refresh the MCP tools.

OAuth installation

Use OAuth when the MCP client supports browser-based authorization and the connection should act as the signed-in Siesta AI user.

1. In Siesta AI, open Profile > Apps & Extensions.
2. Find MCP Server, select the OAuth tab, and copy the configuration. The OAuth definition contains the MCP server URL without API key headers.
3. Add the definition to your MCP client and start the connection from that client.
4. If you are not signed in, Siesta AI opens the login page. Sign in with the account that should authorize the client.
5. On Review connection, verify the client name and host before continuing.
6. Review the requested permissions, then select Allow access or Cancel.
7. Siesta AI returns the authorization result to the MCP client. Continue in that client and confirm that the Siesta tools are available.

The consent screen can request these scopes:

Scope	Meaning
mcp	Lets the client access and use the Siesta AI tools and data exposed through MCP under the signed-in user's access.
offline_access	Lets a compatible client stay connected when the user is not actively using it. This scope appears only when the client requests it.

Only approve a request that you started from a client you recognize. Use Use another account when the displayed Siesta AI account is not the one that should own the connection.

If you select Cancel, the denial is returned to the client and no authorization is granted. If the request is invalid, expired, or interrupted, return to the MCP client and start the connection again. Do not copy authorization codes, request values, or browser callback URLs into client configuration.

Two different MCP directions

This page connects an external AI client to Siesta AI as an MCP server. To connect a third-party MCP server as a tool inside Siesta AI, use the MCP Connection guide instead.

Tools

Once connected, the client can use the tools below. With API-key authentication they operate within the supplied organization; with OAuth they operate under the signed-in user's organization and effective access. The server's advertised tool list remains authoritative for the connected environment.

Agents

Tool	Description
list_agents	List the agents (chatbots) in your organization. Use it to find an agent's ID before messaging or editing it.
get_agent	Read an agent's full configuration: system message, model, connection, skills, sub-agents and tools.
create_agent	Create a new agent from a name and system message, optionally attaching a model, skills, sub-agents and tools.
update_agent	Update an existing agent. Only the fields you pass change; list fields replace the whole set.

Conversations

Tool	Description
send_message	Send a message to an agent and get its reply. Pass an agentId to start a new conversation or a conversationId to continue one.
list_conversations	List existing conversations (newest first), with optional filtering by agent or title search and paging.
get_conversation	Read the full message history of a conversation before resuming it.
get_tool_call_result	Read the complete available result and arguments for one tool call in a conversation.
get_conversation_reasoning	Read the reasoning summaries exposed for a conversation without requesting hidden chain-of-thought.
get_sub_agent_conversation	Read the conversation created by one recorded sub-agent call.
approve_tool_call	Approve one identified pending tool call and continue the conversation. This tool is marked destructive.
reject_tool_call	Reject one identified pending tool call and prevent it from running.

Skills

Tool	Description
list_skills	List the skills (named instructions plus bound tool functions) available in your organization.
get_skill	Read a skill's full detail: its instructions and bound functions.
create_skill	Create a new skill from a name, instructions and optional bound functions.
update_skill	Update an existing skill. Only the fields you pass change; passing functions replaces the whole set.
list_skill_functions	List the tool functions that can be bound to a skill, returning the IDs to use when creating or updating one.

Workflows

Tool	Description
list_workflows	List the workflows in your organization. Use it to find a workflow's ID before reading or editing it.
get_workflow	Read a workflow's full graph: nodes (with IDs, types and input mappings) and edges.
create_workflow	Create a workflow from a full graph of nodes and edges in one call.
update_workflow	Replace a workflow's graph with the full set of nodes and edges it should have.
describe_node_types	Discover the workflow building blocks (node types, their inputs/outputs and selectable targets) before building a workflow.
list_workflow_functions	Search the tool functions that can back a workflow tool node, filtered by provider or capability.
run_workflow	Start a workflow run. This tool is marked destructive because its nodes can change external state.
list_workflow_runs	List recent runs for a workflow.
get_workflow_run	Read one run and its node outcomes.

Discovery

Tool	Description
list_connections	List the connections (integrations) available as an agent's model provider or as a tool.
list_models	List the LLM models available for an agent, optionally filtered by connection.
list_system_tools	List the built-in system tools (for example web grounding and scraping) that can be enabled on an agent.
list_data_source_collections	List Data collections that can be assigned to an agent.

Graph tools

For Graph concepts, lifecycle, Explorer, Galaxy View, and Agent assignment, start with Graphs.

The Siesta AI MCP server exposes Graph knowledge through lowercase tool names. These are different from the internal Graph_* tools that Siesta AI attaches to an Agent with Graph assignments. External MCP clients should use the advertised lowercase tools and treat their current schemas and annotations as authoritative.

Graph access is evaluated for the identity represented by the connection. Graphs.View and the Graph's resource access policy are both required to read knowledge. Mutations additionally require Graphs.Edit and effective write access. list_graphs returns only accessible Graphs and includes canWrite for each result. Treat a not-found response for inaccessible knowledge as an access boundary; do not probe alternate identifiers.

Tool family

Tool	Purpose	Character
list_graphs	List accessible Graphs and whether each can be edited.	Read-only
search_graph_nodes	Search one Graph by text, node types, date filters, paging, and sort order.	Read-only
get_graph_schema	Read observed node types, relationship types, endpoints, and optionally property observations.	Read-only
get_graph_node	Read one node and its current version.	Read-only
get_graph_edge	Read one relationship and its current version.	Read-only
get_graph_neighborhood	Read one page of direct incoming, outgoing, or bidirectional relationships around a node.	Read-only
get_graph_subgraph	Read connected context up to three hops from a node in both directions.	Read-only
create_graph_node, create_graph_edge	Create knowledge with a server-generated identifier after resolving existing entities or endpoints.	Write, non-idempotent
update_graph_node, update_graph_edge	Replace complete knowledge using optimistic concurrency.	Write, destructive, idempotent
delete_graph_node, delete_graph_edge	Delete knowledge using its current version.	Write, destructive, idempotent

Call list_graphs to resolve a graphId and check canWrite. Before creating a node, call search_graph_nodes so an existing entity can be reused instead of duplicated. The search requires at least one query, node type, or creation/update date filter. Resolve both node IDs before creating a relationship.

Paging and bounded output

list_graphs defaults to 25 results and accepts 1–100. search_graph_nodes defaults to 20 and accepts 1–100; follow offset, returnedCount, totalCount, and moreAvailable to retrieve later pages. get_graph_neighborhood defaults to 50 relationships and accepts 1–100; follow every returned continuationToken when the complete one-hop neighborhood is required.

get_graph_subgraph returns compact context up to three hops from one node and always traverses both directions. It is bounded rather than paged. Use node search to resolve the starting node, neighborhood for exhaustive direct connections, and subgraph for focused multi-hop context.

Results can include outputTruncated, omittedItemCount, sourceTruncated, sourceTruncationReasons, or other paging metadata. Treat these fields as part of the result contract. Never infer that a visible response is complete when its metadata says otherwise; the limit reported by the connected environment is authoritative.

Detail levels

Graph read tools use bounded detail levels:

• Summary returns topology, identity, and versions with the smallest output.
• Properties adds properties and can restrict them with propertyKeys.
• Full adds available evidence and returns all properties within the output bounds.

Start with Summary, then request Properties or Full only for relevant nodes and relationships. list_graphs uses its own Summary or Full detail choice, where Full adds bounded descriptions.

Safe write sequence

1. List Graphs and confirm the target and canWrite value.
2. Search for every referenced entity before creating anything.
3. Read an existing node or relationship with Full detail and no propertyKeys before replacing or deleting it.
4. Stop if that full read reports outputTruncated; a complete safe replacement cannot be constructed from truncated content.
5. Send the complete replacement, preserve authorized structured evidence, and include the current expectedVersion.
6. Verify the returned identifier and version before the next mutation.

Create operations are non-idempotent and generate a new identifier. Updates and deletes use optimistic concurrency and fail rather than silently overwrite a newer version. Extraction guidance is configuration, not evidence or authorization to write.

Security metadata and confirmation

Read-only, create, destructive, idempotent, and open-world annotations describe the intended effect of each tool. send_message, run_workflow, updates, deletions, and approve_tool_call are state-changing or can trigger downstream effects. A destructive annotation can prompt warnings in a compatible client, but it does not guarantee a client-side approval dialog.

Before a state-changing call, show the target and complete arguments to the user, apply the client's confirmation policy, and verify the resulting Siesta AI object or external effect. Use the MCP governance matrix for the recommended access mode. Do not describe this endpoint or a client as marketplace-approved unless the publication status has been verified separately.