Ghost Team Brain

agent-discovery-market

9 min read

✏️ Edit this page on GitHub

Agent discovery market & competitive landscape

Transcripts mishear competitor “Ora” as “Aura”/“Auro” — all map to Ora.

Current truth

  • Two named competitors, opposite shapes. Ora (web-readiness tooling, below) and Wani Wani (wani-wani, waniwani.ai — “Make AI your #1 distribution channel”; Granola mishears it as “One-E-One”) — a French-rooted, SF-HQ’d AI-distribution startup building AI storefronts across ChatGPT/Claude/Gemini (insurance/finance-heavy; raised an $8M seed led by Seedcamp, Jun 2026). Wani Wani is a done-for-you build-across-named-channels package, not tooling; the lesson we take from them is packaging (sell the whole distribution outcome), not their tech.
  • The field is widening (2026-06-22). More entrants now visible: metatuner (Rotterdam, post-connection invocation optimization, copied our site), noodle-seed (website↔app conversion layer, “Halo”), Lixo, and two YC companies in the latest batch doing agent discovery — Scope flagged as the one to watch (good launch video). These YC entrants are framed as Ora’s competitors, not ours (web/AEO-shaped). Google Gemini Enterprise released an agent registry/catalog (traceability/discoverability/“AI catalog”, domain-trust based) — significant overlap with what Ora is building, and could undercut Ora’s Cloudflare/Vercel partnerships (a development the founders would welcome).
  • Two discovery layers (the market’s real split, sharpened 2026-06-22): (1) organic / pre-connection — appearing in the connection picker before the user has connected you; only measurable by live client prompt runs (invisible in host logs) — our core, hard to replicate; (2) post-connection — optimizing invocation/tool-calls/ranking once connected, from live data — where MetaTuner and Alpic’s intent→invocation feature play. Most rivals do (2); (1) is the moat.
  • Moat framing (2026-06-22): every hosting platform (alpic, manufact) is incentivized to add a “discovery tab” to keep apps on-platform — but our market is vast (any company, incl. the many not hosted on these platforms) and organic-discovery measurement is a multi-week build, not a feature toggle. The play: become the embedded, host-agnostic discovery layer before platforms build their own, and integrate with all of them. Also: no canonical optimization rules exist (models change; Anthropic doesn’t publish golden tips) → optimization must be model-scoped (e.g. Haiku), which favors a continuous-measurement tool over a static playbook.
  • Two distinct agent worlds framed (2026-06-08): (1) registry-based agents inside the big harnesses (ChatGPT, Claude, Gemini — curated connectors/apps/plugins) vs (2) open-web agents (coding agents, Slack/Teams bots, company-built agents like Salesforce Agentforce, GojiBerry — they search/act on the web, not registries). No public data on the volume split (consumer harness vs API-driven agents) — acknowledged unknown.
  • Ora is the main competitor reference, positioned opposite us: “agent-ready web” across all surfaces, 40+ optimization checks, heavily developer-focused (default test harness Claude Code). Their stat: 92 of the top 100 websites score zero on agent readiness. Founder is ex-Tavily head of research, co-created with Liad (confirmed 2026-06-11) — the same Liad the founders identified as their ideal advisor and decided not to chase (“he needs to come to us”, 06-05). They explicitly do NOT believe in our registry/connector niche (“can’t build a billion-dollar company on it” — relayed second-hand).
  • Our contrarian position: connectors/registries are the more lucrative path over the next 1–2 years (1B+ users in these harnesses; Claude observably prefers connectors over web when one fits). Web optimization beyond basic AEO is unvalidated for ChatGPT/Claude — treat web-readiness as a free marketing-funnel offering, not the core product.
  • Layered offering model sketched: (1) AEO/“tags”, (2) connectors (client-specific capabilities — our core), (3) “everything else” on the agent-ready web (llms.txt, auth.md, agents.json etc.) — relevance grows as harnesses open up (Codex, Cowork use web + tools freely; ChatGPT/Claude are restricted).
  • Top research priority agreed: harness capability mapping — per harness, per prompt type, what surfaces/options does it actually use, and when does it fall back from connector to web/AEO? This defines what’s worth optimizing and is “the golden bullet” for BD.
  • Customer segmentation lens agreed: company/vertical → user persona → which agent platform those users prompt in → use case; companies self-define their persona/platform inputs (we can’t divine that data), which also reveals their AI strategy assumptions.
  • Hierarchy belief: labs prefer connectors (trust, data access); if no connector fits, fallback is web. Both labs becoming “operating systems”; marketplaces for agents AND apps expected; agents will autonomously pick low-risk tools (PDF-tool selection, Statista-style 50¢ micropayments via x402 — Tavily already integrated).
  • Confirmed first-hand by OpenAI staff (Codex Ambassadors office hours, 2026-06-11; elliot-garreffa attended): a merged ChatGPT+Codex experience (media’s “super app”) is coming — both brands remain, ~90–95% of current Codex UI stays, gradual rollout. Validates Elliot’s 06-09 merge prediction. Codex repositioning from developer-first to “for everyone” (financial analysis, slides, busy work) — framing: “every task can be phrased as a coding task”.
  • Codex plugins (same source): connectors currently auth via the user’s ChatGPT account; a unification workstream is in progress; plugin sharing is coming to consumer plans (today team/enterprise only — security review is the delay). Elliot pushed the marketplace framing (shared plugins + ratings = discovery surface) and asked directly whether ChatGPT apps and Codex plugins become ONE store — no answer yet. One OpenAI staffer’s stated personal view of the end state: user types one line, backend picks the tools — i.e. discovery moves under the hood, which is exactly our optimization thesis.
  • The category-education gap is the live BD friction (2026-06-23). In outreach, prospects largely don’t know the connector-discovery problem exists and aren’t thinking about it — they have to be deep in the space to get it, so every sale doubles as category creation (the same “create the market” problem Wani Wani faces). Segmentation read: most outreach so far has targeted the connector-audience segment (companies that already have an app/connector); an untested segment is companies that haven’t started with MCP at all — the plan is to test the full-service “get present” offering on them. Posture: stay confident, it’s a growing market. Near-term revenue lead = direct-booker (gives BD a concrete thing to pitch), with the discovery flow as the parallel growth lever.
  • Field signals (2026-06-16, Awaze call + internal debrief): Kiwi.com reportedly ~500k–1M Claude sessions/mo — validation that connector discovery moves real volume. Big travel brands are deliberately waiting: Airbnb built no connector (a co-founder thinks the ChatGPT app UX isn’t rich enough yet); Marriott built one but is waiting for maturity. Claude enforces a hard-coded rule that connectors can’t directly sell products (no checkout/ecommerce connector). Per one EU operator’s analytics (awaze), Mistral & Perplexity outrank Claude as referral sources (European-audience effect), while GPT-bot crawl volume is large (~1.12M/30d) but converts ~nothing to human sessions. Full research agenda → research-priorities.

Open questions

  • API vs consumer usage split across the agent market — no public data; can we estimate?
  • How will Google let businesses be discovered by Gemini? (Watch for an “MCP apps with 50 launch partners”-style announcement — medium likelihood, reserve capacity.)
  • Why does everyone else (Ora et al.) bet on the open web while we don’t? We could not articulate a confident answer on 2026-06-08 — explicit gap to close with research.
  • Does calling a raw API even allow tool use without a harness? (Vincent doubts it — verify.)
  • “Ash” — unrecognized harness/platform option seen in Ora’s tool — what is it?
  • “Hassan” (garbled, 06-08) — a third Ora person, or mishearing? Liad confirmed 2026-06-11.
  • Will ChatGPT apps and Codex plugins merge into a single store/marketplace? Asked OpenAI directly (06-11): unification workstream exists, no answer given. Matters for whether our ChatGPT-apps and Codex-plugins tracking converge.

Timeline

  • [2026-06-23] (internal discussion — Vincent + Elliot standup) BD-friction read: prospects don’t know about / aren’t thinking about the connector-discovery problem (a category-creation sell); flagged the not-yet-on-MCP segment as untested vs the connector-audience focus so far, to be tested with the full-service “get present” offering. direct-booker reaffirmed as the near-term revenue lead; the full-customer-journey / MCP-distribution play (per 2026-06-15-offering-pricing-fundraise) reaffirmed as the growth engine — explicitly modeled on Wani Wani (Granola again heard “one new one”), who are verticalizing for insurance. Wani Wani headcount/strategy update logged in wani-wani.
  • [2026-06-22] (internal discussion — Vincent + Elliot, website/strategy session) Widened the competitive map: added metatuner, noodle-seed, Lixo, and YC’s Scope (Ora’s competitors); noted Google Gemini Enterprise’s agent registry/catalog overlapping Ora. Sharpened the two discovery layers (organic/pre-connection = our moat; post-connection = others) and the embedded host-agnostic discovery layer moat framing. Logged the no-canonical-rules / model-scoped optimization insight.
  • [2026-06-16] (call + internal) awaze discovery call + internal debrief — field signals above (Kiwi ~500k Claude sessions/mo as validation; Airbnb/Marriott deliberately waiting; Claude no-checkout rule; Mistral/Perplexity > Claude for an EU base). Research agenda set in research-priorities.
  • [2026-06-15] (internal discussion — Vincent + Elliot retro) Reviewed Wani Wani (transcript “One-E-One”) after their $8M seed (led by Seedcamp; verified 2026-06-16) — see wani-wani. Takeaways: packaging is the differentiator they nailed (sell the whole AI-distribution outcome end-to-end, named-channel hook, funnel thinking); their one-step bet (build the apps) vs our two-steps-removed (optimization/distribution) bet — ours more defensible but later-stage. Reinforced “we have to create the market ourselves” framing.
  • [2026-06-11] (community event — OpenAI Codex Ambassadors office hours, ~130 attendees incl. OpenAI staff; Elliot attended) OpenAI confirmed the merged ChatGPT+Codex experience: both brands stay, ~90–95% of Codex UI stays, gradual rollout, possible early access for ambassadors (no promises). Codex going “for everyone” beyond developers. Plugins: auth via ChatGPT account today, unification workstream running; plugin sharing coming to consumer plans (security review pending) — marketplace dynamics expected; one-store question (ChatGPT apps + Codex plugins) asked by Elliot, unanswered. GPT-next models teased: “sooner than you think, later than you want”, no timeline. OpenAI Global Build Week: announce Jul 7, events Jul 13–21 (community events worldwide, broader-than-tech audience). Org/relationship intel: Pauline Navas (ex-Vercel community lead, based Brighton/London) joined OpenAI 06-09 to run the ambassador program; Elliot arranged a London coffee with VB (OpenAI, runs Codex ambassadors) for week of 2026-06-15. ~2h quarterly session on assumptions (“things that need to become true”), predictions/market events (kept as separate lists), and Ora teardown (they live-tested Ora’s tool: it created a Salesforce account autonomously via web crawling; cost Ora ~38¢; positioned as agent-journey debugging for websites). Agreed: focus locked to ChatGPT + Claude (Codex considered as consumer-bound; Gemini watched), don’t boil the ocean; web-readiness score usable as free funnel + advocacy (“you might get an answer, but you’ll never get used without the registry”). Agreed assumptions doc to be finished across multiple sessions during the week (target Fri 2026-06-12). Noted Ora added AEO/GEO definitions seemingly after reading our thesis. MCP dev summit cited as highest-value learning activity last quarter; consider attending another in the fall.

Graph View

Backlinks