Argus

Argus watches public news and market feeds, spots stories that are gaining momentum, and proposes a token idea plus a small product that rides the story. A human decides at every step. Argus never launches a token.

Last updated 2026-11-07

How it works

From a news story to a live preview of a small product, in seven steps.

  1. Read. Argus reads public news, trend and market feeds all day.
  2. Group. It groups related items into a Narrative: one story that is gaining momentum.
  3. Score. An AI judge scores each Narrative on momentum, novelty, buildability, hook and legitimacy.
  4. Propose. For the best ones, argus proposes a token idea plus a small product (the Tech) that rides the story.
  5. Decide. A human, the Operator, reviews the Proposals in the War Room and approves or dismisses them.
  6. Build. For an approved Proposal, an AI agent builds a small working product inside a locked box.
  7. Preview and launch. The product appears as a preview link. Humans decide whether to launch the token, outside argus.

Highlighted steps are where a person decides.

Small glossary

Signal
One raw item seen in a feed: an article, a trending term, a market event.
Narrative
A cluster of recent Signals about one subject that is gaining momentum.
Proposal
A Narrative plus a token angle (name, ticker) and a Tech worth building. What the Operator decides on.
Tech
The small product argus can build that rides the Narrative.
Spec
The written brief for a Build. The Operator edits it before approving.
Build
The small working product made from an approved Proposal's Spec.
Operator
The one person who approves each gate.
War Room
The local dashboard where the Operator reviews Proposals.

Architecture

The path from public feeds to a live preview. Left to right, then down.

Argus architecture Sources feed ingest, then Narratives, the Judge and Propose. Proposals go to the War Room where the Operator decides, then a Spec, a Build in containment and a workers.dev preview. Humans launch in newpairspam, outside argus. SourcesRSS news, Google Trends,PumpPortal, DexScreener,GitHub trendinglive IngestPlain code, no AI,every 5 minutes.Items become Signalslive NarrativesSignals clusteredby meaninglive JudgeAI scores each Narrative.Below the buildabilitygate, no Proposallive ProposeName, ticker, Tech.Checks name clashes.Max 30 a daylive War RoomRuns on localhost.Operator shortlists,approves or dismisseslive, human decides SpecDrafted when shortlisted.Operator edits it,then approvesbuilt, not used yet Buildin containmentAn agent writes theproduct. One at a time,one retrybuilt, not used yet PreviewLive link on workers.devto the finished productbuilt, not used yet newpairspamHumans launch here.Out of scope forargusoutside argus Locked box: rootless podman, no keys inside Also in the picture. No arrows: they touch several steps, or come later. ContainmentRootless podman, no keysinside, egress allowlistdone Launch handoffOpens the newpairspamform prefilleddone Team notificationsChat messages, notify only,no commandsbuilt, not used yet Paid API fallbackBackup for judge and propose.Off by default, $0.50 a day capbuilt, not used yet P2-17 market evidence on each cardPumpPortal activity. Waits for the 48 hour run Phase 3: feedback loopDismissal reasons tune the Rubric; optional auto-approve
  1. SourcesRSS news, Google Trends, PumpPortal, DexScreener, GitHub trendinglive
  2. IngestPlain code, no AI, every 5 minutes. Items become Signalslive
  3. NarrativesSignals clustered by meaninglive
  4. JudgeAI scores each Narrative; below the buildability gate, no Proposallive
  5. ProposeName, ticker, Tech. Checks name clashes. Max 30 a daylive
  6. War RoomRuns on localhost. Operator shortlists, approves or dismisseslive, human decides
  7. SpecDrafted when shortlisted. Operator edits it, then approvesbuilt, not used yet
  8. Build in containmentAn agent writes the product inside a locked box. One at a time, one retrybuilt, not used yet
  9. PreviewLive link on workers.dev to the finished productbuilt, not used yet
  10. newpairspamHumans launch here. Out of scope for argusoutside argus
  11. ContainmentRootless podman, no keys inside, egress allowlistdone
  12. Launch handoffOpens the newpairspam form prefilleddone
  13. Team notificationsChat messages, notify onlybuilt, not used yet
  14. Paid API fallbackBackup for judge and propose. Off by defaultbuilt, not used yet
  15. P2-17 market evidencePumpPortal activity on each card. Waits for the 48 hour runnot built
  16. Phase 3: feedback loopDismissal reasons tune the Rubric; optional auto-approvenot built

Guardrails

What keeps this safe: four rules, plus one.

Human in the loop

A person decides at every gate: shortlist, approve the Spec, dismiss, launch. Argus never launches a token and holds no wallet.

Containment

Builds run in a locked container: no passwords or keys inside, internet only to two allowed sites, no access to the Operator's files. There is no uncontained mode.

Budget caps

Every AI run has a spend cap. Research is held to about 30% of the daily plan. A Build gets a $25 cap and 4 hours. Hitting a cap stops the run; caps are never raised silently.

Subscription only

AI work runs on the Claude subscription, not pay-per-use. The optional paid fallback is off, and capped at $0.50 a day if ever turned on.

Honesty

Every Build carries a footer saying argus made it and the creator may sell. Argus does no posting, botting or promotion.

Where we stand

Status as of 2026-11-07.

Phase 1: build path

Built, not used for real yet

Tickets 01 to 20 are done, including containment. Open: the first real Build (ticket 10) and one check in ticket 12. Both need the Operator present and the Build token. No real Build has run yet.

Phase 2: research

Live in the acceptance run

All tickets are done except three: P2-14 (the 48 hour acceptance run), P2-17 (PumpPortal market evidence, which waits for P2-14) and ticket 20 (in progress). Also shipped: Activity page, Atom feeds, the new War Room look, and Proposals sort and filter (live). Ticket 20 adds a confirmation dialog before Shortlist and an Undo shortlist. Faster pre-push checks (skip for docs-only changes, skip code already checked) are in progress.

The 48 hour acceptance run

Running

Research only, paid fallback off. It stopped once when the computer went to sleep; it was restarted and now holds a sleep/lid lock while it runs. The War Room can now restart without stopping the worker. Ends about 2026-11-08 13:00 (-03:00). Still to check: the forced-exhaustion fallback and the path from card to launched. Usage so far is roughly $0.3 to $0.9 of estimated plan usage per 6 hours (not billed).

Phase 3: feedback and auto-ladder

Not started

Dismissal reasons feed the Rubric; optional auto-approve above a score, off by default. Auto-launch is out of scope.

Waiting on the Operator

  1. Mint the Build token. It blocks the first real Build.
  2. Confirm the workers.dev subdomain in Cloudflare.
  3. Check the extra-usage setting on the Claude account.
  4. After the run: decide the API fallback and the research quota.
  5. Re-check that headless Claude stays on the subscription.

Discussion

Post your answers where the team talks. The Operator collects them. Nothing on this page records replies.

What should we build next?

  • P2-17: show PumpPortal market activity on each card. Worth it? It waits for the 48 hour run.
  • Run the first real Build end to end. Which Proposal first?
  • Proposals list sort and filter is live. Is the order useful? What is missing?
  • Phase 3: let dismissal reasons tune the Rubric.
  • Phase 3: should argus ever auto-approve above a score?
  • More sources: X through the official API, later.
  • Small hardening fixes (timeouts, one podman setting). Now or later?

Questions

  • Is the current Claude plan enough for research plus Builds?
  • Should the paid API fallback ever be on? At what daily cap?
  • Will headless Claude keep drawing from the subscription? Unresolved.
  • Are Narratives good enough? The clustering margin is thin.
  • What should a Team member see on a Proposal to give a useful opinion?

Risks

  • Research and Builds share the plan limits with the Operator's own use.
  • Look-alike token names: the clash check can answer "unknown", not only yes or no.
  • Real Builds are untested: cost and failure rate are unknown.
  • The run stopped once when the computer slept; it now holds a sleep lock.
  • An AI-built product may overpromise. The disclosure footer helps.