Wispear

Wispear is the trust-scored discovery engine for AI agent components, built on Intuition Protocol. AI agents don't have an intelligence problem — they have a discovery, composition, and trust problem. Hundreds of tools, MCPs, skills, and packages exist, but there's no intelligent way to find, compare, or assemble them into a coherent stack. Current signals (GitHub stars, download counts) are noisy and unverifiable.

Wispear changes what the AI receives before it responds. When a user describes their need, the Wispear Agent extracts semantic claims, queries a community-curated knowledge graph where real practitioners stake $TRUST on components they validate, and composes an executable blueprint — not a list, but a plan: what to use, in what order, and why. The system serves three actor levels: casual users get personalized blueprints with zero friction (no wallet needed), connected users save stacks on-chain, and curators put skin in the game by staking on tools within their declared expertise domains.

Key components include an adaptive swipe-based profiling flow, a chat-to-blueprint interface with live visual construction, a component explorer with real-time on chain data.

Every interaction enriches the graph. Usage validates attestations, trust signals sharpen, and recommendations improve continuously through a dual flywheel between users and curators.

Wispear turns builders' collective experience into AI context — so every request benefits from the intelligence of those who already solved the problem.

Wispear is built as a TypeScript monorepo (Bun + Turbo) with 3 apps and 7 packages.

Chat App (Next.js 14) The main discovery interface. A user describes what agent they want to build, and a 4-stage pipeline kicks in:

Extract semantic claims from the intent via Claude Opus Query the Intuition knowledge graph through a custom MCP server to find ranked components (skills, MCPs, APIs, LLMs) Build context with trust scores from on-chain curator votes Generate a full agent blueprint (component stack, system prompt, install commands, MCP configs, flow diagram) via Claude

Curator App (Next.js 14) Lets users vote on-chain for the best components per use-case. Votes go through Intuition Protocol's MultiVault bonding curves on the Intuition L3 chain (ID 1155, native token TRUST). We use nested triples — (component, is-best-of, type) wrapped in ((triple), in-context-of, useCase) — so curation is contextual. Wagmi + Viem + Dynamic Labs handle wallet interactions and batch deposits.

Swipe App (Vite + React 18) A mobile-first PWA with a Tinder-style UX. Users swipe through max 8 adaptive questions (binary decision tree in static JSON) to detect their Role and AI Maturity level, then publish their profile on-chain as Intuition atoms + triples via ethers.js v6 and @dynamic/appkit.

Agent Package Uses the MCP SDK as a client to talk to our Intuition MCP server (HTTP + stdio transports). Notable hack: Intuition's MCP server rejects concurrent connections, so all knowledge graph queries are serialized. Blueprint data is base64-encoded and passed via URL params between apps.

Feedback API (Hono + Drizzle + SQLite) Closes the loop: logs sessions, messages, and blueprint selections so curators can improve rankings over time.