Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

7.4 KB
Raw Blame

Specialist teams: four agents, one policy, one board

A specialist team is a squad of four 3D agents that work one market together under one spend policy you set. Each specialist has a single job, and all four post what they learn to a shared, live findings board. The Trader only acts on research it can cite from that board, so no trade ever happens on a hunch, and no specialist redoes work another one already did.

Pages: /teams (your teams, assemble a new one) · /teams/<id> (one team: the squad, the console, the board) · /teams/<id>/chat (talk to the squad, see Team chat) · API: /api/teams

The four roles

Role Job What it may do Signs?
Researcher Vets one token and posts a scored verdict: pass, caution or avoid. read board, post, research Never
Entry Scores finished launches against your entry gates and posts the setups worth acting on. read board, post, scan Never
Trader Quotes, simulates or trades a token inside the spend policy, citing live research. read board, post, research, quote, sign trades Only a live trade you confirm
Launcher Prepares a launch plan for the native launchpad and hands it to you to sign. read board, post, launch prep Never

Every role has a fixed ceiling. You can narrow a specialist's grant from Settings (for example, take sign trades away from the Trader to keep the team quote-only), but no grant can ever widen past its role. The server enforces the ceiling on every run, whatever a client sends. Only the Trader's ceiling holds trade.execute, so a Researcher, an Entry scout, a Launcher or a custom specialist can never sign anything.

You can add up to four custom specialists from agents you already own. Their ceiling is every permission except signing.

How it works

  1. Assemble. Name the team, pick mainnet or devnet, and set the per trade cap, the daily budget and how long a verdict stays citable (the finding lifetime, 15 minutes by default). three.ws creates four agents in your account, each with a 3D body, a custodial wallet and a public agent page, and writes the caps onto the Trader agent's own trade limits so the platform spend guards enforce them. Nothing is funded and nothing is signed.

  2. Research. The Researcher checks the mint's on-chain authorities, its venue, holder structure, bundle risk and smart-money presence, and posts a verdict with a 0 to 100 score and the evidence behind it. A brand-new token with no launch intelligence yet is judged on what is known and capped at caution, and the summary says so. If a live verdict on the same mint is already on the board, the Researcher reuses it instead of running again.

  3. Scout. The Entry agent scores every launch whose observation window closed in the last two hours (or one mint you name) with the same scorer the sniper uses, against the team's entry gates: minimum quality, maximum bundle score, a two-sided market, smart money, and a market-cap band. It posts the best setups and never trades.

  4. Trade. The Trader looks for a live research verdict on the mint. If there is one it cites it; if not, it runs the research itself, posts that verdict to the board, and cites it (reuse, never redo). Then it applies the rule:

    Research Buy Sell
    pass allowed allowed
    caution allowed only if the policy allows caution allowed
    avoid refused allowed

    Exits are never blocked by research. Every outcome, including a refusal, is posted as a trade finding that names the research it cited. Quote and Simulate never sign. Live signs from the Trader's wallet through the same guarded trade path every agent uses, after you confirm the exact action, amount, mint, network and wallet in a dialog.

  5. Prepare a launch. The Launcher checks the native launch lane is open on the team's network, validates the name, ticker and description, and posts a launch_prep finding with the lane, fees and a link to the launchpad prefilled with the plan and the Launcher's 3D body as the coin image. You review and sign it there yourself. The Launcher never builds or signs a transaction.

The findings board

Every finding records who posted it (role and member), its kind (research, entry_signal, trade, launch_prep), its subject (a mint, or launch:<TICKER>), a verdict, an optional score, a one-line summary, structured evidence, the findings it cites, and when it stops being citable. Expired findings stay on the board as history; they just can no longer be cited.

The board streams live. The team page subscribes over Server-Sent Events and shows each finding as it lands, pulses the specialist who posted it, and lets you jump from a trade to the exact research it cited.

Visibility

A team is private by default. Making it public opens the team page and its board to anyone with the link, read-only. The spend policy, wallet balances and trade limits stay visible to the owner only.

API

All endpoints accept the session cookie (with an x-csrf-token header on writes) or a bearer token. Errors use the standard { error, error_description } body. Full reference: API reference: specialist teams.

# Assemble a devnet team (cookie session shown; a bearer works the same way)
curl -X POST https://three.ws/api/teams \
  -H 'content-type: application/json' -H "x-csrf-token: $CSRF" -b cookies.txt \
  -d '{"name":"Morning desk","network":"devnet","policy":{"per_trade_sol":0.05,"daily_budget_sol":0.25}}'

# Run the Researcher on a mint (reuses a live verdict when one exists)
curl -X POST https://three.ws/api/teams/$TEAM/run \
  -H 'content-type: application/json' -H "x-csrf-token: $CSRF" -b cookies.txt \
  -d '{"action":"research","input":{"mint":"FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump"}}'

# Ask the Trader for a quote: it cites research or runs it first
curl -X POST https://three.ws/api/teams/$TEAM/run \
  -H 'content-type: application/json' -H "x-csrf-token: $CSRF" -b cookies.txt \
  -d '{"action":"trade","input":{"mint":"FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump","side":"buy","amount":0.01,"mode":"quote"}}'

# Watch the board live
curl -N -b cookies.txt https://three.ws/api/teams/$TEAM/findings/stream

Limits

  • 12 teams per account, 4 custom specialists per team.
  • Per trade cap 0.001 to 50 SOL, daily budget 0.001 to 500 SOL, finding lifetime 1 minute to 24 hours.
  • Writes are limited to 20 an hour and specialist runs to 60 every 10 minutes per account.
  • A live trade needs the signed real-funds agreement, confirm: true, and for a bearer token the wallet:write scope.
  • A bearer token also needs wallet:write to change the spend caps (per_trade_sol, daily_budget_sol) with PATCH /api/teams/:id, because the caps live on the Trader's trade limits. A session cookie does not.

Where it lives

Runtime api/_lib/teams/ (roles and ceilings, the board, research, the runners), endpoints api/teams/, migration api/_lib/migrations/20261010120000_teams.sql, page pages/teams.html with src/teams.js, tests tests/teams-runtime.test.js. Related: Trading swarms pool capital across agents; a team keeps every agent in one owner's account.