AI orchestration engine
Tholo, an AI orchestrator that pulls only the skills a task needs.
It turns a large skill catalog into a focused build, pulling only what a task needs.
What is Tholo?
Tholo is an open-source AI orchestrator for the Claude Skills catalog, built by RampStack. It reads the catalog, plans a build task, and selects only the skills the task needs, then prepares the work to hand to Claude Code.
The idea
Select, do not load everything.
A skill catalog that grows past a few dozen entries develops two failures. Skills start to clash, more than one looks applicable and they pull the work in different directions. And loading the whole catalog to do one task floods the context with skills the task will never use.
Tholo addresses both by selecting.
When a task is too broad to narrow, Tholo says so rather than dumping everything. When nothing fits, it says that rather than forcing a pick.
What Tholo does
Five jobs, each shaped by the catalog and the stop.
Reads the catalog.
Enumerates the skills, their descriptions, categories, and the when-to-use and when-NOT-to-use boundaries each skill declares.
Plans a task.
Decomposes a build task into steps, each a coherent unit of work the catalog has skills for.
Selects per step.
Scores the catalog against each step and picks only the relevant skills, with the inputs each skill needs already gathered into the dispatch.
De-conflicts overlaps.
Uses each skill's own stated boundaries to resolve clashes when more than one skill could apply, so the dispatch reflects the catalog's own declared lanes.
Stops where it should.
When a step is too broad to narrow, Tholo asks rather than guesses. And it never runs the build itself, since running writes files and changes a system.
The stop
It plans, you run.
Running a build writes files and changes a system, which is irreversible, so it is the human's call to trigger. The honest stop is the boundary.
How Tholo works.
Tholo is the AI orchestration engine. Give it a goal and it selects the catalog skills that address it, resolves the overlaps between them, sequences them into steps, and produces a prepared dispatch: each step with one goal, the skills it draws on, and the inputs it needs before it can run.
Selection is the hard part, and it is where the catalog's own structure does the work. Every skill in the catalog declares what it is for and when not to use it. Tholo scores candidates against the goal, and when two skills claim the same ground, the declarations decide it: a skill that names another in its when-not-to-use gives Tholo the seam between them. The result is a small set of skills that fit the task, not a pile of everything plausibly related.
The output is a dispatch, prepared and not run. Steps that have what they need are marked ready. Steps that are missing an input say so, and the dispatch names the missing piece instead of guessing it. A person reviews the plan, supplies what is missing, and approves it before anything executes.
The pipeline
Select a stage to see what happens there.
Goal. One goal in plain language is the input, the outcome the build should reach.
A worked example, in two parts.
This is Tholo running live today. When a Basano audit finds failures, the verdict is handed to Tholo, and Tholo prepares the fix.
Input: a Basano verdict for a product page with five failures: missing image alt text, a missing canonical link, thin structured data, AI crawlers blocked in robots.txt, and no llms.txt at the site root.
What Tholo prepares
Step 1
ReadyAccessibility
Skills:accessibility-auditmedia-asset-management
Fix the alt-text failure at its source, the page and the asset pipeline behind it.
Step 2
ReadyTechnical SEO and AI search readiness
Skills:seo-technicalseo-aeo-geo
Add the canonical, deepen the structured data, allow AI-crawler access, publish llms.txt.
The selection seams show in what is absent. The catalog has fourteen SEO skills; the dispatch names two. seo-onpage scored as a candidate for the alt-text fix and was kept on the accessibility step's periphery rather than added as a third SEO skill, because accessibility-audit already owns the failure. The orchestration suite, the backlink skills, and the diagnosis skills never enter, because nothing in the verdict calls for them.
Output: a prepared dispatch a person can read in one pass, approve, and hand to whoever does the work, an agent or a person.
The same engine, given a plain goal instead of a verdict. A goal can carry more than one phase, and Tholo splits it into ordered steps before it selects anything.
Input: "Audit and fix the SEO on an existing page."
What Tholo prepares
The decomposition
One compound goal, split at the seam into two ordered steps. Select a step to read what it covers.
Step 1: Audit. The first phase, taken from the left of the coordinating word. Tholo prepares it as its own step.
The split is the judgment here. Tholo reads the goal, finds the seam at the coordinating word, and prepares an audit step and a fix step on the same page, in that order. It does not invent phases the goal does not name: a goal with a single phase stays one step, and a goal too broad to split stops for a person to narrow. Each step then runs through the same selection shown in the first tab.
Output: two ordered steps, audit then fix, each prepared and not run.
What the two parts show together.
The same engine on two different kinds of input. A verdict gives Tholo verified failures, and it maps each to the skills that fix it. A plain goal gives Tholo an outcome to reach, and it splits the goal into ordered steps before selecting per step. In both cases the output is a plan a person approves, with every skill named, every step scoped to one goal, and every gap stated instead of papered over. The catalog supplies the skills; Tholo supplies the judgment about which few, and where the steps divide.
What it does
One task at a time.
Each card is one task Tholo handles: what goes in, that Tholo runs it, and the artifact that comes out.
Tholo workflows
Pick a workflow to see what goes in, and the artifact that comes out.
One goal in, a prepared dispatch of the few skills it needs out.
A goal. One goal in plain language, like build a landing page for a new product.
Read the code. Read the catalog.
The repo is the working implementation and the worked examples. The catalog is what Tholo runs on. Solutions is how to bring RampStack in to operate it for you.