Skip to main content

​
Skills and plugins

  • A skill is a packaged set of instructions that an agent loads when a task matches it, on top of whatever tools it already has. Skills follow the open Agent Skills specification.
  • A plugin is an installable bundle of skills for Claude Code, registered through a marketplace so a single command installs the whole set.
Allium ships two plugins from the allium-labs/skills repo: allium-agent (direct data access via the allium CLI and partner infrastructure skills) and allium-analyst (the analyst methodology covered on this page). Install either one on its own.

​
The allium-analyst plugin

Match the right source

Your agent compares candidate tables, dashboards, and endpoints against the question before writing SQL.

Build the query and check it

It checks Allium Terminal for existing coverage, builds from a small test query to a production one, and validates the numbers.

Sector-specific pitfalls

It checks for the known double-counting, classification, and aggregation mistakes in RWA, stablecoins, bridges, DEX, and lending.

Reproducible results

It saves load-bearing queries to Explorer so anyone can rerun them.

​
Installation

You need the Allium MCP already configured before installing this plugin. See the Quick start if you haven’t set that up yet.
1

Add the Allium Labs marketplace

In Claude Code, run:
2

Install the plugin

3

Restart Claude Code

Restart so the new skills load.
4

Try it

Ask something that needs the workflow, for example:
Use the Allium skills to investigate how tokenized RWA treasury supply has grown over the past 6 months, and save the query so I can share it.
It’s wired up correctly if your agent checks Terminal for existing coverage and hands you back a saved Explorer query at the end.
If you’re not on Claude Code, or want specific skills without the plugin system, use the Agent Skills installer:
Add --skill <name> (for example --skill data-matching) to install one skill. Without it you get every skill in the repo.

​
Skills reference

data-matching and allium-investigation are the workflow skills. The rest are sector skills. Your agent loads one once the question touches that sector, and it plugs into the workflow at the source-selection step.
It defines a data contract for the question (object, measure, grain, scope, and time semantics), compares candidate sources against that contract, and proves coverage with a bounded test query: a chain or table that appears in search may hold no real data. It runs first, before allium-investigation and any sector skill.
It runs the investigation on the source data-matching picked. It checks Allium Terminal for existing coverage, builds the query from a small ephemeral test into a saved production query, and validates the result: reconciliation, duplicate-count checks, and cross-checking at least one figure a different way. It labels each finding as observed fact, interpretation, or unknown. At the end it saves load-bearing queries to Explorer, and when the investigation produced several queries it offers a dashboard hand-off to Allium Assistant.
It covers treasuries, money market funds, tokenized equities, commodities, and private credit. It treats AUM and supply as a snapshot and never sums them across days, aggregates multi-chain fund deployments and discloses the aggregation, classifies by underlying risk factor and not by legal wrapper, and excludes known bad-data tokens before ranking holders.
It covers USDC, USDT, DAI, and other issuers. It separates raw volume from adjusted volume, where raw can overstate activity by 5 to 20 times, handles chain-specific double-counting such as XRPL’s trust-line rippling and Aptos’s dual token addresses, and applies the address case-sensitivity rules, which vary by chain.
It covers CCTP, LayerZero, Wormhole, Across, and canonical or application bridges. It explains the asymmetric, single-directional way transfers are recorded, the double-counting that shows up when you filter to two or more target chains at once, and how to trace a lock-and-mint transfer across chains when no direct join key links the two legs.
It covers Uniswap, PancakeSwap, Pump.fun, Orca, Pendle, aggregators, and perp DEXs like Hyperliquid. It explains per-hop double-counting on multi-hop routes, when to group by protocol and when to group by project, and how to reconstruct an aggregator’s implicit fee when it arrives as a plain transfer with no labeled column.
It covers Aave, Morpho, Fluid, Euler, Maple, and their forks. Deposits, withdrawals, loans, repayments, and liquidations live in separate event tables, so it joins them to detect operations with no dedicated label, such as collateral swaps and deleveraging. It also handles the column-naming inconsistencies across those tables and fill-level double-counting on liquidations.

MCP Overview

Set up the Allium MCP server itself

Tools Reference

Documentation for the MCP tools these skills call

allium-labs/skills

Skill source, README, and the allium-agent plugin

Examples

Things customers built with the MCP