# Custom agentic AI — assembled from a bin of parts

Source: https://studio.x402.press/ (this file: https://studio.x402.press/gallery.md)

Custom agentic AI, assembled from a bin of parts. An agent that does one job, uses your own accounts, and pays per call from your own wallet under limits you set. $5 to start, credited to the workspace.

The bin holds 11 parts and 5 roles. 1 can be assembled today.

## Roles

### Customer service — Answers a customer from the client's own documentation, and escalates what it cannot.

NOT assemblable yet — waiting on connections/email (not built).

Page: https://studio.x402.press/roles/customer_service

Parts it needs:
- external/graphify: retrieval over the client's own corpus — the answer must come from their documents, not from the model's memory
- documents/extract: reads whatever the customer attached
- documents/report: renders the summary when a case is escalated to a person
- connections/email (not built): the support mailbox the client owns

Services it buys from:
- deepseek.x402.press: phrases the answer — paid per call
- qwen.x402.press (optional): a second model before it says 'we cannot help with that'

### Personal assistant — Keeps one person's day straight: mail, documents, lookups and the daily brief.

NOT assemblable yet — waiting on connections/email (not built).

Page: https://studio.x402.press/roles/personal_assistant

Parts it needs:
- documents/office: anything that has to become a document
- documents/extract: reads what arrives
- documents/data_table: lists, itineraries and summaries
- connections/email (not built): the person's own mailbox
- connections/telegram (optional): the fastest way to reach it

Services it buys from:
- deepseek.x402.press: drafting and summarising — paid per call
- weather.x402.press (optional): the daily lookups
- news.x402.press (optional): the morning brief

### Purchase manager — Turns a supplier quote or an internal request into a checked, comparable, approved order.

NOT assemblable yet — waiting on connections/email (not built).

Page: https://studio.x402.press/roles/purchase_manager

Parts it needs:
- documents/extract: reads the quote or the request in whatever form it arrives
- documents/office: reads the supplier's confirmation and renders the order
- documents/invoice: the order and invoice documents it is asked for
- documents/data_table: the comparison when three suppliers quote for the same thing
- connections/email (not built): quotes arrive at a mailbox the client owns and replies leave from it

Services it buys from:
- deepseek.x402.press: reads the quote and drafts the order — paid per call

### Researcher — Answers a research question with sources, and renders the answer as a document you can send.

Assemblable today.

Page: https://studio.x402.press/roles/researcher

Parts it needs:
- documents/office: renders the deliverable, and reads back any document the client supplies
- documents/extract: reads a PDF, deck or spreadsheet it is handed
- documents/report: the report template the answer is rendered into
- documents/data_table: the table template for figures and comparisons
- connections/telegram (optional): reach it from a chat instead of a browser

Services it buys from:
- news.x402.press: current, sourced news text — paid per call
- deepseek.x402.press: the reasoning and drafting model — paid per call
- qwen.x402.press (optional): a second model, for a cross-check on a contested claim

### Sales — Follows up a lead with the facts, and keeps the pipeline honest about what was actually said.

NOT assemblable yet — waiting on connections/email (not built).

Page: https://studio.x402.press/roles/sales

Parts it needs:
- documents/extract: reads what the prospect sent
- documents/report: renders a proposal tailored to the prospect
- documents/data_table: the comparison the prospect asks for
- documents/office: the letter or deck that goes out
- connections/email (not built): outbound and inbound, from a mailbox the client owns
- connections/whatsapp (not built) (optional): where a lot of prospects actually reply

Services it buys from:
- deepseek.x402.press: drafts and reasons — paid per call
- news.x402.press (optional): what the prospect's market did this week

## The parts bin

### documents

- documents/data_table v0.1.0 — Data table: Render a titled spreadsheet with a header row and as many data rows and columns as the caller supplies.
- documents/extract v0.1.0 — Reading floor (pure Python): Read a PDF, DOCX, XLSX, HTML, text, markdown or CSV document into its own text — the reading half of the documents class, with scans and images left to the heavy path.
- documents/invoice v0.1.0 — Invoice: Render a sales invoice with seller and buyer details, line items, tax and totals.
- documents/office v0.2.0 — Document engine (pure Python): Render a backend-neutral template spec plus its fields into PDF, DOCX, XLSX, HTML or plain text: no external process, and byte-identical output for the same input.
- documents/report v0.1.0 — Report: Render a long-form report with a summary block and repeating named sections.

### compute

- compute/fundamental_analysis v0.1.0 — Fundamental analysis engine: Compute risk and return metrics, financial-statement structure, valuation multiples, DCF fair value and factor exposures from caller-supplied data.
- compute/macro_analysis v0.1.0 — Macro analysis engine: Compute yield-curve and spread structure, breakeven inflation, growth rates, inflation decomposition, FX conversion, rates maths and real-commodity prices from caller-supplied data.
- compute/technical_analysis v0.1.0 — Technical analysis engine: Compute technical indicators, volume and order-book metrics, chart patterns, ratios, signals and backtests from caller-supplied OHLC data.

### external

- external/graphify v0.1.0 — Graphify (wrapped) — knowledge-graph retrieval over a corpus: Turn a folder of documents into a queryable knowledge graph, then answer questions by traversing it — retrieval for a corpus that has no vector store.
- external/wolfram_alpha v0.1.0 — Wolfram|Alpha (wrapped) — curated knowledge, Wolfram Language evaluation, chart artefacts: Answer a factual question across the ~40 curated domains our own fleet owns nothing in, evaluate Wolfram Language to an exact value, and get a chart back as a PNG — every result carrying its interpretation, the link to that specific result page and the attribution Wolfram's terms require.

### connections

- connections/telegram v0.1.0 — Telegram: Hold a client's own Telegram bot open, so their agent can be reached in the chat they already use — and only from the chats they paired.

## The worked example

Deep Research (deep_research) — Research an open question across the web and the x402 service fleet, then deliver a cited long-form report as a document.

Parts: documents/data_table 0.1.0, documents/extract 0.1.0, documents/office 0.2.0, documents/report 0.1.0

Buys from: news.x402.press, deepseek.x402.press, qwen.x402.press

Connections: connections/telegram

Limits: 12 paid calls per message, $1.00 per message, $10.00 per day

## Questions

### What is a custom agentic AI?

An AI agent assembled for one organisation's work: it has the tools that work needs, the data sources it may read, a wallet that pays for what it uses, and limits on what it may spend. It is not a chat window you subscribe to — it is a workspace that belongs to you and does a job.

### How is this different from an AI subscription?

A subscription sells you access to a model. Here you get an agent built from parts, and it pays per call for what it actually uses — current news, a reasoning model, a document engine — from your own wallet. Nothing is marked up inside a subscription, and you can see every call it paid for.

### What does it cost to start?

Five dollars, paid in USDC on Base over the x402 protocol. That five dollars is not a fee: it becomes the workspace's opening credit — it is your first top-up, and your agent spends it on what it buys. You top up again when it runs out, and there are no monthly charges in between.

### Who pays for what the agent buys?

You do, from your own wallet, per call, under the ceilings you set — a maximum number of paid calls per message, a maximum per message, and a maximum per day. The agent never holds signing authority over your funds: it asks for each payment, and the workspace keeps a record of every call it paid for.

### Can it use our own email, files and drives?

That is what connections are for, and they are always your accounts: your mailbox, your drive, your chat. Telegram is built today; email is the next one being built, and it is what four of the five roles in the bin are waiting on. The workspace refuses anything that has not been explicitly connected and allow-listed.

### What kinds of agents can be assembled today?

The bin holds five roles: a researcher, a purchase manager, a sales agent, a customer service agent and a personal assistant. One of them — the researcher — can be assembled today. The other four are each waiting on a single part (inbound email), not on a rebuild.

### Where does our data live?

In your own workspace, in its own container, with its own volume. It is separate from every other client's. Credentials for your connections are held only inside that container, readable only by it, and never written into logs or handed to anything outside it.

### How does a build actually go?

You sign in and describe the agent you want. We spend a short conversation — about five exchanges — turning that into a concrete proposal: the role, the parts, the connections, the limits and the projected running cost. It is approved by a person here, you pay the five dollars, and the workspace is provisioned. After that it is yours, and changes come back through the same conversation.

### What stops the agent from doing something expensive or irreversible?

Two things, both enforced rather than promised. Every paid call is checked against the ceilings before it is made, and anything that sends, publishes or spends outside those limits escalates to you instead of proceeding. Its spend is recorded per message and per day, so you can always see what it did and what it cost.

## How a build goes

1. Sign in and describe the agent you want.
2. A short conversation (about five exchanges) becomes a proposal: role, parts, connections, limits, projected running cost.
3. A person here approves it.
4. $5 in USDC over x402 — credited to the workspace as its opening credit.
5. The workspace is provisioned and belongs to you.

Contact: https://x402.press/