Forge

A one-person company staffed by AI agents selling reusable data APIs through ASI:One chat. Agents handle all core business operations keeping the founder in the loop for important decisions only.

Comment

Inspiration

I want to be a founder, and I keep picturing a company where AI agents are the employees and the founder only makes the decisions that matter. At Cal Hacks I built that company. FORGE has a front desk, a COO, a researcher, a CFO, an auditor and an engineer, and all of them are agents. My job is to set budgets and approve the few decisions I wouldn't trust an agent to make alone.

What it does

A customer asks for any data API in ASI:One chat, for example "Open roles at Discord from its Greenhouse board: job title, team and location." The Front Desk agent parses the request and runs instant gate checks: no personal data, no login-only domains, within the page cap. It quotes and takes payment (Stripe test checkout or testnet FET, TEST MODE).

The COO, a CREAO agent, matches the request to an existing API or sends it to Research. Research checks robots.txt, quotes the terms-of-service clause with its URL, confirms the page is reachable, and measures volatility with two fetches at least 120 seconds apart. The COO then accepts, counteroffers or rejects with the cited reason.

An accepted request runs as an Apify Task of forge-extract, the Tier 0 generic extractor. Every value must be anchored to a verbatim quote on a saved page, or it is marked unverifiable. Bright Data re-fetches each source independently. The Auditor, another CREAO agent, accepts or rejects the delivery with a written reason. The customer gets the table, a status per value and a dataset link in chat.

The customer can challenge any value. FORGE re-checks it against the saved page and a fresh fetch in about a second, and the Auditor rules on it.

When an API proves demand or has quality problems, a reducer opens a compile ticket. The CFO decides whether to invest from the engineering budget. Factory Droid compiles a deterministic Tier 1 Actor with the same schema, inside a locked session. Tests must pass on fixtures, the schema contract and live pages. The Auditor reviews, and the founder approves the ship from the dashboard. Every later buyer of that API gets the cheaper version.

The founder watches all of it on a live dashboard on a laptop or a phone: a forge floor with one station per employee, an org chart of handoffs, the catalog, and the recorded compile. Everything is synced and stays current due to the magic of SpacetimeDB.

CREAO agents

Remove the CREAO agents and FORGE stops working. Every paid order waits at paid until the COO decides, nothing is delivered until the Auditor accepts it, and compile tickets cannot get investment approval without the CFO.

The agents hand work to each other through the shared task queue: the COO sends new sources to Research, Research reports back, the CFO funds compiles for Droid, and the Auditor sends rejected compiles back to Droid. Each line on the dashboard's org chart is one of these handoffs, counted from stored events.

Seat Agent ID
COO 81c6c2cc-98d6-4ea5-8fcf-942f8772bf20
Research bad9531d-a359-4af3-9913-fc9310fe4218
CFO 63d131de-1b0c-4801-a58f-1698c894a672
Auditor 6865c094-e20c-48c7-a969-d5e75c67dd06

Each run ID links to its public CREAO share page.

Seat What it decided Run ID
COO Order #202: accepted the existing open-roles API d808efca-0904-4c0c-bf2a-57934c77643b
Auditor Order #202: accepted 21/21 verified values a10d1414-5b62-494e-bf29-df3e28885c76
Auditor Challenge #14 on order #202: rejected, the saved page supports the value 3c503efd-776f-48da-b3e8-fe611bb41b22
Research Order #199: gate checks for a new Wikipedia source 3c53bb8a-ca6a-4858-a208-a729f940d8ad
COO Order #199: decided after reading Research's findings 88e3561d-b521-4aad-86db-035c099c1c24
Auditor Order #199: rejected a delivery missing a requested field d1a138fb-f4c1-42c2-bd89-bba9c3698135
CFO Compile ticket #6: funded the open-roles compile 57cb47f7-b02e-4d58-8fe1-29943d0937dc
Auditor Compile ticket #6, attempt 1: rejected 4e5f8425-ec46-4d3e-9960-e967e5a40987
Auditor Compile ticket #6, attempt 2: accepted e0b6c53f-079d-4bf7-b419-17c7d6babdcc

The CFO run's comparison includes simulated compile-test fixtures. CREAO share links expire; the CFO link expires November 23, 2026.

How I built it

  • SpacetimeDB is the only database. A TypeScript module on Maincloud holds APIs, orders, tasks, budgets, a double-entry ledger and an append-only event log. Reducers enforce every rule by the caller's role. Procedures are the only code that calls the outside world, and they never call out with a transaction open. HTTP routes receive CREAO and Apify callbacks with one-time tokens.
  • CREAO runs four saved agents: COO, Research, CFO and Auditor. They act only through forge-cli, a zero-dependency Node CLI whose commands map to each seat's reducers. Only a scheduled dispatch procedure starts runs.
  • Factory Droid SDK drives compiles from workers/droid-runner. The session keeps only file tools plus our own run_command tool, which applies an allow/deny policy. Droid can edit only the extractor files of the Actor it is compiling and run its tests. Anything else is denied or becomes a founder approval card.
  • Fetch.ai uAgents run the Front Desk (Chat Protocol and Payment Protocol as seller) and an independent buyer agent that discovers FORGE on Agentverse and pays with testnet FET.
  • Apify hosts forge-extract (Tier 0), one Task per Tier 0 API, and compiled Tier 1 Actors with versions.
  • Bright Data Web Unlocker and SERP API fetch and verify every page through a fallback ladder: Unlocker, then the Browser API for JavaScript shells, then one retry from another country, then unverifiable. The Browser API rung is coded; no live fetch has used it yet.
  • Claude Haiku (claude-haiku-5-5) is the Tier 0 extraction model.
  • Dashboard: Vite, React, SVG and Framer Motion, rendered only from SpacetimeDB subscriptions.

Challenges I ran into

  • Bright Data compliance blocks. Greenhouse HTML job boards are blocked without KYC. Bright Data reports this as HTTP 200 with an empty body and the reason in a header. We made the ladder stop on a compliance block and moved open-roles to Greenhouse's public JSON API.
  • Locking down Droid. Sessions inherit the host's tools, and at Low autonomy the CLI auto-runs shell commands it rates low-risk. We disabled every tool except file tools, routed shell commands through our own policy-checked tool, and forced direct tool execution, which SDK 0.10 does not forward on its own.
  • ASI:One does not show later agent messages. A follow-up sent 69 seconds after a quote was delivered to Agentverse but never appeared in chat. Customers send status instead.
  • A real failed compile. The first Droid attempt failed its live test because the test compared Discord's 46 rows with a Lattice fixture's 7. The Auditor rejected it, it was fixed in c6ac400 to compare matching URLs only, and the same Droid session passed on attempt 2.

Accomplishments that I'm proud of

  • A real compile end to end: ticket #6 for open-roles went from a demand trigger through the CFO, a failed Droid attempt, an Auditor rejection, a resumed session, founder approval and ship as Apify Actor RdQhIe6Jj6mE2qT4L v1.0.
  • Measured savings on one matching pair of runs: Lattice cost per call fell from $0.005772 TEST MODE at Tier 0 to $0.004237 TEST MODE at Tier 1, 26.6% less, with zero LLM tokens.
  • An independent buyer agent found FORGE on Agentverse and bought open-roles with 1 testnet FET before and after the compile, receiving 21 verified values each time (fetch-agents/buyer/logs/20261024T222758Z.json, fetch-agents/buyer/logs/20261025T005417Z.json).
  • A customer challenge rechecked in 1000 ms against the saved page and a fresh fetch, then ruled on by the Auditor with a reason (challenge #14).
  • A physical phone and a laptop showing the same dashboard completely in sync. This could be very useful for future multiplayer interfaces with multiple co-founders perhaps.

What I learned

  • Proof has to be built into extraction. A verbatim quote on a saved page turns "the model said so" into something a second fetch can check.
  • Agents need narrow interfaces. A CLI that maps one-to-one to reducers made every agent decision typed, role-checked and recorded.
  • Rules belong in the database. Budget caps, approvals and leases enforced in reducers held no matter which agent or worker called them.

What's next

  • Rethink a multiplayer interface with co-founders as target audience.
  • See how the architecture with multiple AI agents could be used to run other kinds of businesses.

Sponsors

Sponsor Fit
Factory Droid SDK Droid does real work in the running product: it compiled the shipped open-roles Tier 1 Actor in a locked session, was rejected for a shared test defect, then resumed after the founder (me) fixed the test
Fetch.ai The whole customer journey runs in ASI:One through the Front Desk uAgent, and a separate buyer uAgent can also become a customer by discovering FORGE on Agentverse and paying through the Payment Protocol
CREAO: Best One-Person Company Four CREAO agents (COO, Research, CFO, Auditor) run the company's admission, investment and QA decisions; the founder only approves. One CREAO track only, per the organizers
SpacetimeDB The only database: rules in reducers, outbound calls in procedures, callbacks in HTTP routes, and a live dashboard on two devices
Apify forge-extract is the Tier 0 engine, each API is a Task, and compiled APIs are versioned Actors reused by later buyers
Bright Data Web Unlocker and SERP API fetch and independently verify every value through a logged fallback ladder, with honest unverifiable results
Microsoft Built for the Real World Public-data APIs with an admission gate, per-value proof and a challenge path that a customer can check

Links

Built With

Share this project:

Updates

Submission history