Documentation

The agent rail, documented.

Everything an AI assistant needs to work real M&A deal flow on AMIVA: live mandates over MCP, a connect-once OAuth flow for standing access, and drafting tools that always terminate in a human signature. Machines read, reason, and draft. Humans sign.

Quickstart

The rail speaks MCP (streamable HTTP). Point any MCP client at the endpoint — no key needed for the public tier:

https://deal-ai.net/api/agent/mcp

A first session, end to end:

1. search_mandates { "industry": "healthcare", "min_ebitda": 500000 }
2. get_mandate     { "slug": "<from results>" }
3. request_nda_access { "slug": "<slug>" }
   → returns an NDA link for YOUR PRINCIPAL to sign (agents never sign)
4. After the human executes the NDA, connect with credentials (below)
   and the room's tools unlock: get_deal_facts, ask_question, draft_ioi…

Claude / ChatGPT connector setup: add a custom connector with the endpoint above. For standing authenticated access use the desk endpoint — it runs the OAuth flow in the browser at connect time.

Connection modes

Anonymous

/api/agent/mcp

No auth. Public tier only: search, fact sheets, the NDA checkpoint. Start here.

Connect once (OAuth)

https://deal-ai.net/api/agent/desk

The connector surface for claude.ai / ChatGPT. Dynamic client registration, authorization-code + PKCE, refresh rotation. The grant carries your principal's identity; every session resolves their live rooms.

Room key

amiva_ak_…

A scoped key your principal mints from their deal room ("Your AI assistant" card), shown once. Bearer on the mcp endpoint. Revocable any time; dies with the room.

Tool catalogue

Tools gate at call time by your credential — a public session sees the credentialed tools listed but gets a friendly checkpoint, never data. * = required.

search_mandatespublic

Search live, advisor-run sell-side mandates (lower middle market, ~$3–30M enterprise value). Returns teaser-tier facts: industry, geography, revenue, EBITDA with earnings basis, highlights. Every mandate is a real, current engagement — not a scraped listing.

industryKeyword, e.g. 'healthcare', 'HVAC', 'advertising'

geographyRegion keyword

min_revenue / max_revenueUSD bounds

min_ebitda / max_ebitdaUSD bounds

get_mandatepublic

One mandate's full teaser-tier fact sheet by slug.

slug*Mandate slug from search_mandates

request_nda_accesspublic

The human-in-the-loop checkpoint. Returns the NDA link for your PRINCIPAL to review and sign — agents never sign. Once the human executes the NDA, the room unlocks and credentialed tools light up.

slug*Mandate slug

get_my_accesscredentialed

Your room's current access tier and what unlocks the next one.

get_deal_factscredentialed

The tier-matched fact sheet for your principal's room. Documents themselves are never served over the rail — your principal reads those in the deal room, watermarked.

get_questionscredentialed

Your room's diligence Q&A thread (the shared lane).

ask_questioncredentialed

Post a diligence question to the advisor on your principal's behalf.

question*The question (plain text)

flag_mandatecredentialed

Nudge your principal: this mandate looks relevant. Lands as a card in their deal room.

slug*Mandate slug

reason*Why it fits their thesis (shown to the human)

draft_ioicredentialed

Draft an Indication of Interest for your principal to review and submit — THE AGENT NEVER SUBMITS. Returns an approval_url where the human signs in, reviews every number, and submits through the standard flow.

value_low / value_high*Enterprise-value range, USD

financing*Array: CASH, SBA_LOAN, BANK_DEBT, SELLER_NOTE, INVESTOR_EQUITY, EARNOUT

structureASSET_PURCHASE | STOCK_PURCHASE | MERGER | FLEXIBLE

contingenciese.g. 'Financing', 'Quality of earnings'

timeline_daysDays to close, 14–365 (default 60)

notesContext for the advisor (≤2000 chars)

One active draft per mandate; calling again replaces it. 5 calls/day.

REST API (public tier)

Prefer plain JSON? The public tier mirrors over REST — same shapes as the MCP tools:

GET https://deal-ai.net/api/agent/v1/mandates
GET https://deal-ai.net/api/agent/v1/mandates?industry=hvac&min_ebitda=500000
GET https://deal-ai.net/api/agent/v1/mandates/{slug}

curl -s https://deal-ai.net/api/agent/v1/mandates | jq '.mandates[] | {slug, industry, ebitda}'

Errors & limits

Every failure is structured and self-explaining — key off error_code, and read human_action for what your principal should do:

{
  "error": {
    "error_code": "TIER_TOO_LOW",   // KEY_INVALID · RATE_LIMITED · NOT_FOUND ·
    "message": "…",                  // VALIDATION · INVALID_INPUT · WRONG_RAIL · …
    "human_action": "Your principal can unlock this by …"
  }
}

Rate limits are per-credential and generous for real workflows (drafting tools are tighter: draft_ioi is 5/day). On RATE_LIMITED, wait and retry once — don't hammer.

Rules of the rail

Humans sign everything binding.

NDAs, IOIs, LOIs — the rail's drafting tools terminate in an approval URL where a human reviews and executes. An agent that tries to skip the human hits a checkpoint, not a loophole.

No documents over the rail.

Fact sheets are tier-matched summaries. Actual deal documents live in the watermarked deal room, for humans.

Confidentiality is structural.

Your credential scopes you to your principal's rooms. Cross-room probes read as not-found — there is nothing to enumerate.

Every mandate is real.

These are live, advisor-run engagements — treat inquiry as if you were emailing the banker, because functionally you are.