Build with Hum

MCP servers

Tool servers give Hum abilities beyond the machine. The ones your account comes with connect at sign-in. Add your own in a few lines.

Included with your account

Servers your account is granted connect when you sign in and at the start of every session. Nothing to configure. A server granted later appears the next time you start hum.

Every account has docs: Hum's own documentation as a tool server. Ask Hum what it can do or how a setting works and it reads the page before answering. For a reader outside a session the same pages are an index at /llms.txt and markdown at /docs/<page>.md.

/mcp              # every server: what mounted, and what did not, with the reason
/tools            # every tool the model has, and the sandbox it runs in

Your own

# ~/.hum/config.toml
[[mcp]]
name = "jira"
url = "https://.../mcp"

[[mcp]]
name = "db"
command = "npx"
args = ["db-mcp", "--readonly"]
timeout_s = 900              # seconds one call may take; default 900

Each [[mcp]] table mounts into every session. /mcp shows whether it mounted.

Key	Description
`name`	What the server's tools are called by (`name__tool`)
`url`	Streamable-HTTP address
`command`, `args`	Start the server here instead
`headers`	Sent with every HTTP request
`timeout_s`	Seconds one call may take before it fails; default 900

What a server gets

A server that asks for a file or a directory as base64 gets it from Hum. The model names a path under the working directory; Hum packs and encodes it.
A server runs with the reach of its own process. The sandbox does not extend to it.
Credential-like variables (*KEY*, *PASSWORD*, *SECRET*, *TOKEN*, *CREDENTIAL*, HUM_*) are stripped from a server's environment. Pass the ones it needs:

[sandbox]
env_pass = ["JIRA_TOKEN"]

Sub-agents

A sub-agent calls the same servers the session has. See Sub-agents.