Agents
Tools and MCP servers
The built-in tools every LoCrew agent gets, such as send_message, and how to connect local and remote MCP servers, approve their launch, and grant tools to agents one by one.
Agents use tools in two ways: LoCrew's own built-in tools, which let them talk to the conversation and each other, and tools from MCP servers you connect and grant.
#Built-in tools
Every Claude Code, Codex and model agent (if the model accepts tools) can call LoCrew's own tools. They are served by a small gateway inside LoCrew that listens only on 127.0.0.1. Each run gets its own token, created when the run starts, held in memory and retired when it ends, so an agent cannot pretend to be another agent by writing a different name.
| Tool | What it does |
|---|---|
send_message | Posts a message to the conversation and, if it names agents in to_agent_ids, wakes them. This is the only way one agent hands work to another. See Agent hand-offs. |
read_messages | Reads recent messages (1 to 100, default 25). |
list_channel_members | Lists everyone in the conversation with their agent ids. |
get_channel_context | Topic, open tasks, remaining agent-to-agent hops and whether hand-offs are allowed. |
get_agent_status | Another agent's runtime, status and whether it is busy. |
update_task | Updates a task the agent is assigned to. Requires Can update tasks it is assigned. Marking a task completed records the agent's claim; it is not proof the work is correct. |
External A2A agents do not get these tools.
#Connect an MCP server
Open Settings → MCP Servers and press Add MCP server. Connecting a server grants nothing: you choose which agents may use which tools afterwards.
#Local server (stdio)
A program started on this computer.
| Field | Notes |
|---|---|
| Name | Unique. |
| Command | Paste a full command line; it is split into the command and its arguments. Example: npx -y @modelcontextprotocol/server-filesystem. |
| Arguments | One per line. |
| Working directory (optional) | |
| Environment variables | Values are encrypted and never shown again. Put credentials here. |
The first time it starts, and whenever the command changes, LoCrew shows a native Launch local MCP server dialog with the exact command, working directory and environment variable names. It warns you when the command asks for elevated privileges, contains a destructive file command, downloads and runs a script, runs an encoded command, or downloads a package the first time it runs. Choose Launch only if you trust it.
#Remote server (Streamable HTTP)
| Field | Notes |
|---|---|
| URL | For example https://example.com/mcp. |
| Authentication | None, Bearer token or Custom header. |
| Token, Custom headers | The token is stored encrypted. Custom headers are encrypted when you mark them secret or their name looks like a credential. |
| Allow plain http:// to a non-local host (not recommended) | Off by default. |
#Both kinds
Request timeout (seconds) is 1 to 600, default 30. Connect on startup connects it when LoCrew starts; it never re-prompts, so a local server whose command you have not approved waits for you to press Connect. Press Add and connect.
#The server card
A card shows the transport, status (Connected, Connecting, Disconnected, Error), the command or URL, how many tools it offers and which agents use them. Test reconnects and rediscovers: Connection works. Found N tools, N resources, N prompts. Expand Tools, resources and capabilities to browse them. A tool's risk label (read-only, writes, destructive) is what the server says about itself; LoCrew does not verify it.
#Grant tools to an agent
In Create agent → Tools, or the agent editor's MCP tools section:
- Each connected server lists its tools. Tick a tool to grant it, or use Grant all / Clear.
- A granted tool is Ask first by default: every call opens the approval dialog. Switch it to Allow to let it run without asking.
Tools appear to the agent as server__tool. Grants are checked on every call, so revoking one takes effect immediately. Removing a server removes its grants.
Note
Work sessions never cover MCP tools. An Ask first tool always asks, even while an agent has temporary write access.
#Errors
| Where | Message | What to do |
|---|---|---|
| Server card | Launching this server needs your approval. Click Connect to review the command. | Press Connect and approve. |
| Server card | Launch was not approved. | You chose Cancel. Press Connect again. |
| Server card | Timed out while connecting. | Raise the timeout or check the command. |
| Server card | The server closed the connection. Last output: … | Read the output in the Server tab's Recent output (stderr). |
| Agent run | The MCP server "…" is not connected… Ask the operator to reconnect it in Settings → MCP Servers. | LoCrew tries one quiet reconnect first. Reconnect it. |
| Agent run | You are not permitted to use "…", or it is no longer available. | Grant the tool to that agent. |
| Agent run | You denied this operation. | You (or a closed dialog) denied an Ask first tool call. |
Not supported yet: OAuth for remote servers, servers that only speak the newest protocol revision, and the SSE and WebSocket transports.