The Mammoth agent skill

mammoth skill install copies a packaged skill—authentication, discovery, structured input, job handling and confirmations—into your coding agent's own skills directory, so it reads the playbook once rather than re-deriving CLI conventions from trial and error.

mammoth skill install
mammoth skill list

What it contains

The skill ships as one SKILL.md entry point plus a references/ directory the entry point routes into by task, rather than one long document read in full every time.

Top-level references:

• about-mammoth.md—what Mammoth is, its data model, and which command does a goal stated in business terms ("combine these files"), including each web-app Transform-menu task
• auth.md—signing in, profiles, and authorized scope
• capabilities.md and capabilities-misc.md—which routes are proven, known-blocked or untried on the release endpoint
• command-index.md—a lookup for one exact command, not upfront reading
• handoff.md—the checklist for pausing or transferring work to another agent
• input.md—nested input documents and how to shape a request
• jobs-drafts.md—waiting on asynchronous work and draft-mode pipelines
• machine-output.md—the envelope shape, in full
• operations.md—which id is which, appending rows, and other one-off rules
• recovery.md—failure modes by message, and the stop-condition list
• report-checklist.md—what to verify before stating a number or calling a step done
• retention.md—deliverable retention and cleanup authorization
• safety.md—confirmation policies and how to satisfy each one non-interactively
• task-start.md—cold start, task planning, and what to do when a route is unsupported

Plus two subdirectories: commands/ (one file per command family, e.g. view.md, dashboard.md, automation.md) and recipes/ (worked flows: end-to-end.md, transforms.md, exports.md, dashboards.md, scheduling.md, database-import.md, cleanup.md, auth-scope.md, resources.md, need-action.md, typed-etl-discovery.md, and an index.md routing between them).

What it gives the agent

• How to authenticate without a terminal, and—just as important—when not to: the skill tells an agent to hand mammoth auth login to the human operator rather than run it, or ask for a key in chat, itself.
• How to discover a command's request shape with mammoth schema get, rather than guessing it from a --help string.
• How to pass structured input as one --input document instead of many flags.
• How to wait on jobs and navigate draft-mode pipelines.
• How confirmations work, and which commands demand which of them—see blast radius and safety.

It is the same material as the agents section as a whole, delivered where the agent will actually read it: inside its own skills directory, at the start of a session.

Choosing targets and scope

The installer knows three agents—Codex, Claude Code and Cursor—and two scopes. The default is user scope, across all three:

Agent	Default destination (user scope)
Claude Code	~/.claude/skills/mammoth-cli/
Codex	~/.agents/skills/mammoth-cli/
Cursor	~/.cursor/skills/mammoth-cli/

Project scope installs into the same relative paths under the nearest directory containing a .git folder instead of the home directory. Review the destinations before a project-scoped install, and target one agent at a time if you do not want all three:

mammoth skill path --output json --no-input                              # show source + every computed target
mammoth skill install --input '{"agents": ["codex"], "scope": "project"}'

mammoth skill show prints the bundled SKILL.md, or one reference file by path, without installing anything—useful for reading the playbook before deciding to install it:

mammoth skill show
mammoth skill show --input '{"file": "references/recipes/end-to-end.md"}'

There is also mammoth skill agents-md install, which writes (or refreshes) a short steering block into a repository's AGENTS.md pointing back at this skill, so a later session in the same repository starts here without re-running the installer.

Keeping it current

Run mammoth skill update after a CLI upgrade. It refreshes copies the installer owns and reports modified copies instead of silently overwriting them, so local edits are never clobbered without telling you:

mammoth skill list      # are installed copies present and intact?
mammoth skill update    # refresh the ones the installer owns

mammoth skill uninstall removes installer-owned destinations; a copy you modified by hand is left in place rather than deleted.

Looking for MCP?

The skill is the CLI's own agent interface. For an MCP client that acts on a workspace, use the Mammoth MCP server. To look up Mammoth documentation from Claude, connect the documentation MCP server: see Use Mammoth with Claude.

---

Next up: For coding agents for install, credentials and the operation loop · Blast radius and safety for what each command demands before it runs.