Native MCP server

Give your agent a human to ask.

Connect Claude Code, Codex, Cursor, or any Streamable HTTP client to Actionbox. A person answers in the hosted inbox; the agent resumes with a durable typed result.

Checking session… Read the MCP guide

ACTIONBOX / MCPconnected

AGENT ASKSask_human

YOU ANSWERWeb dashboard

WORK CONTINUESget_action → response

Secure runner setup

One prompt. Every MCP host.

The prompt contains no Source key. Local runners use the protected CLI bridge; remote runners inject a masked environment secret.

Raw axb_live_… keys must never enter chat, logs, command arguments, or version control.

MCP security and authorization

Recommended local setup

actionbox configure
actionbox doctor

# MCP host command
actionbox mcp

Universal setup prompt

Set up Actionbox as a remote MCP server for this project.

Prefer the local Actionbox CLI bridge. Run actionbox configure outside chat, then actionbox doctor, and configure this MCP host to launch: actionbox mcp. Never paste or print the raw Source key.

If a local command cannot be launched, connect to https://api.actionbox.cloud/mcp using the literal environment reference:
Authorization: Bearer ${ACTIONBOX_SOURCE_KEY}

Keep ACTIONBOX_SOURCE_KEY in the runner's masked secret store. Never run env, printenv, or echo on it; never put it in chat, logs, command arguments, project files, or git.

After configuration, verify with tools/list, create one safe boolean Action with ask_human, then read it back with get_action.

Streamable HTTP endpoint

https://api.actionbox.cloud/mcp

Exact tool inventory

ask_human, get_action, list_actions, update_action, cancel_action, and resolve_action. All six are Source-scoped; machine cancellation and resolution are never attributed to a human.

Typed human input

Boolean, choice, text, numeric, rating, and form interactions.

Source-scoped identity

Every runner only sees the Actions created by its own Source.

Durable evidence

Read the status and typed response after a person answers.