Talk notes
Agent-native sourcing is here, not coming. Funds that haven't yet built agent workflows are at the same disadvantage that funds without spreadsheets were at in 1985. The reason agents change the sourcing math is throughput — a Claude agent can screen 4,200 organizations in 90 seconds against a custom thesis prompt, compared to the 8 hours it would take a human analyst to skim the same list. The bottleneck stops being analyst time and starts being signal quality.
We expose the engine through three protocol surfaces: a free MCP server (Model Context Protocol — the standard adopted by Claude, Cursor, Windsurf, and most agent runtimes), a REST API with per-call pricing, and an x402 USDC endpoint for fully autonomous agents that pay-as-they-go in stablecoin. The MCP server is free forever. The REST API has a free tier and a metered tier. The x402 endpoint charges $0.19 USDC per deep-signal call, settled on Base.
Wiring your agent against /api/mcp/rpc takes about 15 lines of code. There are six read-only tools available: get_signal, list_top_movers, get_org, list_sectors, get_changelog, and search_orgs. A typical sourcing-agent prompt loads the MCP server, asks the agent to score the top 50 movers in the user's investment thesis sectors against a custom rubric, and outputs a ranked shortlist with rationale. The whole flow runs under 90 seconds and consumes about $0.50–$1.20 in API credits depending on agent verbosity.
The pricing model matters for agents. Seat-based licensing (€499/month per analyst) is hostile to agent workflows because the agent doesn't have a seat — it has a workload. Per-call pricing (€0.19/call) is agent-native — the cost scales with usage. We argue this is the dominant pricing model for the next decade of agent-driven analysis tools. We launched it in May 2026 and it's already showing the right pricing-elasticity curve.