Available for workUsually replies within one business dayAI agent delivery · custom tools
We put AI agentsinto working products.
Two things are the core: wiring models and agents into real processes, and building the tools specific people actually open every day. Below are the live products, the tools we have made, and how we can work together.
Role
Gholl Studio / small team
Based in
China · UTC+8
Main stack
TypeScript · React · MCP
You get
Source + deployment + docs
Project index15
Click any row for details
01Selected Work
What we have built
The three below are the ones we would present. The fuller index follows, and every row maps to code that runs.
15
projects
M
01Live
MCPTypeScriptNode.js
012026Live
MCP Micro Tools
42 online tools shared by humans and AI agents
A set of interactive micro-tools on mcp.gholl.com: 42 of them, no sign-in, usable the moment you open one. It doubles as the tool layer for agents — the same capability can be clicked by a person or called over MCP by an agent, without maintaining two versions.
42 online micro-tools, no account required
One capability set serving both web users and agents
Exposed over the MCP protocol so agents can call it directly
A data modelling and query console with role permissions
dbfactory.l.gholl.com keeps data sources, modelling and querying inside one account system. Administrators own accounts and data sources, editors keep building the model, read-only accounts can only query. The layering exists so different roles can safely share a single entry point.
Three account types: admin manages accounts and sources, editor models, viewer queries
Standalone console with username and password login
Modelling and querying happen in the same entry point
Rebuilt a link directory into a personal project index
It used to be a tool directory backed by a cloud database: sites, categories, plugins, favourites and submissions all went through one backend. When that database was retired, the entire dynamic layer died at once. The rewrite collapsed it into a static site — content is code, validated at build time, zero backend at runtime.
Cloud data layer replaced by typed files in the repo, imported at build time
Tailwind and fonts moved from runtime CDNs into the npm build
2,300 lines of admin, login and submission code removed with the backend
SourceReact 19 · TypeScript · Vite 6 · Tailwind v4
Full index
M
01Live
A set of interactive micro-tools on mcp.gholl.com: 42 of them, no sign-in, usable the moment you open one. It doubles as the tool layer for agents — the same capability can be clicked by a person or called over MCP by an agent, without maintaining two versions.
42 online micro-tools, no account required
One capability set serving both web users and agents
Exposed over the MCP protocol so agents can call it directly
dbfactory.l.gholl.com keeps data sources, modelling and querying inside one account system. Administrators own accounts and data sources, editors keep building the model, read-only accounts can only query. The layering exists so different roles can safely share a single entry point.
Three account types: admin manages accounts and sources, editor models, viewer queries
Standalone console with username and password login
Modelling and querying happen in the same entry point
It used to be a tool directory backed by a cloud database: sites, categories, plugins, favourites and submissions all went through one backend. When that database was retired, the entire dynamic layer died at once. The rewrite collapsed it into a static site — content is code, validated at build time, zero backend at runtime.
Cloud data layer replaced by typed files in the repo, imported at build time
Tailwind and fonts moved from runtime CDNs into the npm build
2,300 lines of admin, login and submission code removed with the backend
A plugin that chains ask → retrieve → verify → archive. The model side runs multi-round deep research, produces structured findings and writes them into a Notion database, so nothing has to be tidied by hand afterwards. AI calls and the Notion integration are separate service layers, which makes either one replaceable.
Positions and transactions are organised by account, and every entry can carry the review written that day — so what you did and why you did it stay in one view. A localisation layer sits underneath, with all interface text resolved by key.
High-frequency developer chores collected into one panel: JSON formatting, string encoding, UUID, hashing, colour conversion, case conversion, system info. Everything runs in the browser, so nothing you paste leaves the machine.
Seven tools: JSON, encoding, UUID, hash, colour, case, system info
Computed locally — pasted content never leaves the browser
Most QR sites either generate or scan. This plugin puts both in one interface: the generator takes style options, the reader works from the camera or an uploaded image.
Archives the daily 30-minute state broadcast by date, with day-by-day paging and a digest view. It answers the specific problem of not remembering what aired and not wanting to scrub through video.
The conversation layer abstracts model calls into its own service, fed by the host site’s model registry (priority, temperature, max tokens, health-check thresholds). You can switch vendors on the same question and watch the differences in style and reliability.
A digital implementation of a traditional method: casting, the six-palace reading, and direction calculation, wrapped in visuals serious enough that it works as a tool and a toy at the same time. Built mostly to test whether we can make an irrational brief feel credible.
Built for groups where links get collapsed and unreadable: upload an image to generate share material and create a copyable passcode. Failure paths — upload failed, creation failed — each get their own message instead of one silent error.
A drawing panel with configurable prizes and pools, switching presets per event, with a reveal animation. It exists because after building one for each event, a reusable piece was cheaper than starting over.
The plugin system itself is maintained as a product: one scaffold, two integration modes (iframe and ESM module), agreed size and theme parameters, and a registry with switches and ordering. Adding a tool never requires touching the host site.
iframe and ESM modes behind one container
Scaffold with directory conventions, theme and size parameters ready
Registry switches and ordering without host-site changes
design.gholl.com is an experiment: keep the requirement prompt fixed and swap the DESIGN.md spec, and the product pages an agent produces come out completely different. The site holds 5 in-house baselines and 74 real brand specs, each turned into front-office and back-office showrooms from the same golden prompt. It ships no sample code — only spec data, the prompt, and preview routing.
79 DESIGN.md specs: 5 in-house baselines plus 74 collected real brands
One golden prompt across showrooms, with the difference coming from spec rather than code branches
@google/design.md lint as a gate: every spec must stay at 0 errors
mockoffice.gholl.com is a static capability test for agents: three independent office workflows (expense approval, weekly report rollup, asset request), each made of real pages, with data regenerated every round. The agent reads on page A, carries values across to page B, then submits on page C, leaving a trail only in localStorage, and the home Scorecard rules on it line by line against the evaluation contract in spec.json.
Three office workflows, each a set of operable pages rather than a questionnaire
Data randomised every round, so answers cannot be memorised
spec.json declares the evaluation contract; the home Scorecard rules pass or fail per item
These four are things we have delivered. Other briefs are worth a conversation too — we will say up front whether we are the right team.
AI agent delivery
01
Moving an agent from demo to something you can ship: capabilities split into callable tools, several model vendors with fallback, context and prompt design, results landing in your own systems. We have built a tool service used by people and agents alike.
Deliverables
Capability split and tool interface design
Tool service over MCP / API
Model selection, switching and failure fallback
Write-back integration with existing systems
Good fit · A real process you want an agent to take over part of
Custom tool building
02
Repetitive work collected behind one entry point: internal efficiency tools, data consoles, single-page utilities, batch processing. Scope follows how often something is actually used.
Deliverables
Requirements and interface design
Frontend and backend implementation with permissions
Deployment and usage documentation
Good fit · Something done by hand every week that nobody will greenlight
Data and back-office systems
03
A modelling and query console with an account system: admins, editors and viewers each get exactly the capabilities their role needs, without building a separate product per role.
Deliverables
Account and role permission model
Data source and modelling workflow
Query entry points and read-only views
Good fit · Data that lives long and looks different per person
Product sites and web apps
04
From information architecture to shipped: personal sites, product marketing sites, admin consoles, mobile H5. Designed around the content rather than a template, with a build that does not depend on third-party runtimes.
Deliverables
Information architecture and page design
Frontend implementation, responsive
Build, deployment and domain setup
Good fit · Turning an idea into a site that stays maintainable
Daily toolbox
AI agent orchestrationMCP tool servicesMulti-vendor fallbackTypeScriptReact 19Node.jsNotion API
AI agent orchestrationMCP tool servicesMulti-vendor fallbackTypeScriptReact 19Node.jsNotion API
Vite 6Tailwind CSS v4Data modelling & RBACIndexedDBCanvas motioni18nStatic builds
Vite 6Tailwind CSS v4Data modelling & RBACIndexedDBCanvas motioni18nStatic builds
03How We Work
How a project runs
Scope is fixed in writing before work starts, and you can see a live version throughout.
01
Talk it through
One call covering the goal, the boundaries and what you already have. We come back with a scope judgement rather than an hours estimate.
30 min
02
Plan and quote
A written list of what is in, what is out, the milestones and the cost of each. Scope is locked here; changes go through a variation.
1 – 2 days
03
Build and sync
Each milestone ships a preview URL you can open at any time, so progress is visible instead of landing on the last day.
Per plan
04
Launch and care
Deployment, domain, build scripts and documentation handed over together, with a free fix window afterwards.
30-day window
04Questions
Worth saying up front
All of these have been asked more than once. These are what we actually do today.
01How do you charge?
A fixed price for a defined scope, not an hourly burn. The plan states what is in, what is out, the milestones and the cost of each; later changes go through a variation rather than a quiet renegotiation.
02How big a project can the team take on?
Best fit is anything that ships in one to eight weeks. On larger work we split it into modules and run them in parallel, or take a single module and the technical calls around it; past two or three parallel tracks we flag the schedule risk first.
03Frontend only?
Frontend and backend, with the strength in turning an idea into an interface and a flow people can actually use. Where the problem is really heavy backend architecture, concurrency or data migration, we say so and name what falls outside our range.
04What is the stack?
TypeScript, React, Vite and Node.js; models behind OpenAI-compatible interfaces and MCP; on the data side a modelling and query console with role permissions.
05What happens after launch?
Deployment, domain, build scripts and documentation are handed over together, and anything broken in the following 30 days is fixed free. After that, per-change or an ongoing arrangement.
06Which language do you work in?
Chinese by default, English is fine — this site is bilingual on purpose.
05About
Why this particular pile of things
Gholl Studio started from a very specific annoyance: the tools online were either ad-filled shells somebody else owned, or one website per function. So we built a host site of our own, then turned each scattered need into a small tool behind one protocol — QR codes, JSON formatting, news digests, a trading diary — and shipped them as we went.
The focus later moved to AI agents, but the method did not change: for an agent to be genuinely useful it needs a set of callable tools and a call path you can trust. mcp.gholl.com opens 42 micro-tools to people and agents at the same time; the data factory keeps modelling and querying inside one permissioned console. Models and vendors get replaced — the interfaces and the shape of your data are what stays.
Almost everything here is something we use ourselves, which also makes the bar for building a feature very practical: will it get opened a second time.
How we work
Cut scope until it can ship, then talk architecture
Agent capability has to land on a callable interface, not a demo
No backend where the browser is enough; no open endpoint where roles are needed
Hand over documentation and a reproducible build
How this site is made
A static site built with React, TypeScript and Vite. Content lives in the repository, every page is pre-rendered to HTML at build time, and apart from the ad script there is no third-party runtime dependency. It is the first entry in its own index.
Available for work06 / Contact
Tell us what you are trying to do
Three things are enough to start: the problem, the date you need it live, and a rough budget. The more specific, the faster the plan.