FleetPrompt — The Open Agent Marketplace

Discover, publish, and install production-ready AI agents. Built on trust scores, versioned manifests, and one-click deployment.

The landing page is GENERATED. Do not hand-edit index.html.

/index.html, /replay.js and /form.js are emitted by build-site.mjs from records/surface.json, src/landing.html, src/shell.css, src/replay.js and src/form.js. An edit to the served HTML is silently reverted by the next build. Change the record or the template.

./site.sh     # run the suite(s), emit the site, run the publication gate

build-site.mjs does not quote a test count — it executes the suite, parses what the runner prints, and refuses to emit anything if a number has moved off records/tests.json. That is how the count on this page stopped being a thing anyone could type.

launch-gate.mjs reads the emitted artifact and refuses to publish when it and the records disagree: a retracted claim reinstated, a rung invented, a CTA the rung has not earned, an unrendered token, a mailto: or an email address, a text token below 4.5:1, a same-origin link that resolves to nothing, an artifact that is not what this build emitted, an identifying-animation constant leaking into the copy, or a button whose colour is decided by a non-button rule. It prints its own total — 124 on 2026-08-17, and that figure is the only one in this repository that is not derived, which is why it carries a date. Do not hand-type a check count anywhere: it went from 92 to 124 in a single session, and a typed count is exactly how a printed number and a published one drift apart.

Built against `shell-r9`, recorded as shell_revision in records/surface.json. The whole shell is documented in ProjectAmp2/agents/SHELL.md.

The band says "a specification in the ComputeDriven world", not "the distribution layer of ComputeDriven", and that is deliberate. ampersand-nav records this domain as place: 3, whose renderPlacement() gives the layer sentence to place: 2 only. The old eyebrow claimed the layer and is now blocklisted.

Contact is the Formspree endpoint ruled by Travis (SHELL.md r9), as a real <form> that posts with scripting off; src/form.js only upgrades it to an inline reply and prints success solely on a 2xx. No mailto: anywhere.

Source-of-truth spec

• docs/spec/README.md — FleetPrompt technical specification

• prompts/BUILD.md — Implementation build prompt (9 phases)

Stack

Language:     Elixir 1.17+ / OTP 27
Framework:    Phoenix 1.8 (LiveView for UI, PubSub for trust broadcasts)
Database:     PostgreSQL via shared Supabase (fleet.* schema, migrations 030-039)
Search:       pg_trgm + ts_vector (no external search infra)
Hot Cache:    ETS (:fp_manifests, :fp_trust_scores, :fp_search_index, :fp_categories)
Background:   Oban (trust recompute, webhook delivery, search index refresh)
Auth:         Supabase Auth (shared [&] ecosystem — amp.profiles, amp.workspaces)
Deploy:       Mix release + Docker on Fly.io

Build commands

mix deps.get
mix compile --warnings-as-errors
mix format --check-formatted
mix test                          # requires running Supabase for DB tests
mix test test/fleet_prompt/       # unit tests (no DB required for most)

Key modules

Supabase schema

Tables live in fleet.* schema (migrations 030-039 in ampersand-supabase/migrations/):

• fleet.publishers — marketplace publisher profiles

• fleet.agents — agent listings with search_vector

• fleet.agent_versions — semver-versioned releases with JSONB manifest

• fleet.manifests — typed manifest records (dedicated table, migration 032)

• fleet.trust_scores — computed trust per agent

• fleet.installs — installation records

• fleet.categories — agent taxonomy

• fleet.audit_events — append-only audit log

LiveView pages

• / — SearchLive (⌘K agent search with real-time results)

• /agents/:id — AgentDetailLive (manifest, trust, version history, permissions)

• /publishers — PublisherLive (publisher listing and profiles)

• /trust — TrustDashboardLive (trust scores with tier filtering, PubSub updates)

API endpoints

• GET /api/health — health check

• GET /api/agents/search?q=... — search with min_trust, category, runtime filters

• GET /api/agents/:id — agent detail

• GET /api/agents/:id/manifests — version history

• GET /api/agents/:id/manifests/:version — specific manifest

• POST /api/pipeline/intake — Agentelic ConsolidationEvent intake

Role in [&] Ecosystem

FleetPrompt is the distribution layer and canonical trust broker:

SpecPrompt (spec) → Agentelic (build) → OS-008 (enforce) → FleetPrompt (distribute) → RuneFort (observe)

Key constraints

• Trust scores are computed, never self-reported (4-signal weighted)

• Published versions are immutable (fix via new version)

• Fork trust starts at 0 (forks earn trust independently)

• FleetPrompt does NOT make LLM calls (trust computation is deterministic)

• Install requires explicit permission acceptance

• All tables workspace-scoped with RLS