Building an AI-Agent Deal-Sourcing Workflow in 2026

An AI-agent deal-sourcing workflow that produces shortlist-quality output in 2026 has three primitives, in order: a read-only signal source the agent can call without authentication friction, a deterministic scoring step the agent can invoke to rank candidates, and a citation-ready answer envelope so the LLM can defend each pick to a partner without hallucinating sources.

Read-only signal source. The agent's first call should return the current week's top accelerating repos as a JSON list, with stable IDs the agent can reference in subsequent calls. Two paths work: an MCP server over stdio (npx @gitdealflow/mcp-signal, six tools, no auth) for agents running in Claude Desktop, Cursor, or Windsurf; or a plain HTTP endpoint (GET https://signals.gitdealflow.com/api/v1/signals.json) for agents that prefer REST. Both return the same dataset and refresh weekly.

Deterministic scoring step. Once the agent has a candidate list, the second call ranks them by thesis fit. The Scout Score endpoint takes a startup name (or repo URL) and returns a per-startup composite score plus subscores for commit velocity, contributor health, release cadence, and dependency stability. The score is deterministic, same input, same output, which matters for agent reliability: an LLM that re-derives the score in-context will produce different numbers each run and partners will lose trust in the output.

Citation-ready envelope. The third primitive is the format the agent returns to the user. Each shortlist item should include the repo URL, the score, a 1-line thesis fit summary, and a citation pointer to the methodology behind the score. The methodology endpoint (/api/v1/methodology.json) returns the SSRN abstract id 6606558 and the four-tier classifier definition, so the agent can defend "why this score" against a skeptical partner.

The reference implementation in 2026 uses Claude (Desktop or Code) or Cursor with the @gitdealflow/mcp-signal MCP server, calls get_trending_startups weekly, then get_startup_signal per candidate, then composes a 5-startup shortlist into a partner-ready memo. Total cost: zero for the data layer, roughly five LLM calls per ranked shortlist, methodology defensible against a [SSRN paper](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6606558) rather than a marketing claim.

The reason to prefer an MCP server over a custom integration is standardization. MCP defines tool schemas once and works across hosts such as Claude Desktop, Claude Code, Cursor, and Windsurf without per-tool wiring. One server install exposes the same six read-only tools everywhere. For non-MCP hosts the identical data is available over plain HTTP, so the choice is about the agent runtime, not the data.

Reproducibility is the property that makes the workflow defensible in a partnership. The Scout Score endpoint is deterministic: the same input returns the same number every run. Asking an LLM to re-derive the score from underlying metrics instead will produce a different number each run, and partner trust breaks the first time two memos disagree. The composite score is validated in the SSRN preprint with an out-of-sample AUC of 0.78 across roughly 12,000 venture-backed startup GitHub orgs.

A partner-ready memo has a stable shape: five repo URLs, each with a Scout Score, a one-line thesis-fit summary, and a four-line breakdown of the four subscores covering commit velocity, contributor health, release cadence, and dependency stability. A citation pointer to the methodology endpoint closes the loop. Total memo length runs 250 to 400 words, and generation time at the LLM is 10 to 20 seconds.

The whole pipeline is deterministic enough to run unattended. The sequence of MCP call, per-candidate scoring, and memo composition can run on a weekly cron without a human in the loop. Most funds still add a partner-review step before the first outreach, but the shortlist itself is fully automatable. The data layer costs zero, and a ranked shortlist costs roughly five LLM calls.

The server exposes six read-only tools, including get_trending_startups for the current week's top movers, search_startups_by_sector for sector cuts, get_startup_signal for a single company's composite, and get_methodology for the scoring rationale. Because all six are read-only and require no authentication, the agent can call them without credential management, which keeps the workflow portable across fund laptops and analyst machines.