For agents
GetDeal, callable from your AI agent
GetDeal's deal data, valuation models and AI analysis are available to any AI agent over the Model Context Protocol — Claude, ChatGPT, Cursor, or anything else that speaks MCP. Point your client at one endpoint, present a key you created in the dashboard, and the agent can value a company, read live deal flow, work through diligence and follow your own pipeline without leaving the conversation.
Every answer arrives with its source, the date it was true as of, how it was derived, and an explicit list of anything the caller was not permitted to see.
Capabilities
What your agent can do
Twenty-five tools: twenty-four that answer a question about the market or about your own side of it, plus one diagnostic. Which of them your agent is shown depends on the scopes on the key it holds — a client never sees a tool it would only be refused.
Valuation and benchmarks
Nothing here needs a listing, an NDA or a relationship. It is the same model behind the free valuation tool, with the reference data an agent needs to explain its own answer.
- valuate_startuppublic:read
An indicative valuation range for an AI or software company, from sector revenue multiples scaled by a growth band and a margin band. It needs an annual revenue figure, and it returns a range — the width is the point, and the midpoint is not a price.
- get_valuation_multiplespublic:read
The revenue multiple range GetDeal uses for each sector, with the growth and profitability adjustment factors. Reference data, so an agent can explain or reproduce a range rather than quoting one it cannot account for.
- get_market_snapshotpublic:read
Current activity on GetDeal: how many companies are listed, split by raising versus selling and by funding stage, plus the median amount sought. Money figures are withheld when too few listings match to report them without identifying those listings.
- benchmark_metricspublic:read
Where an ARR figure sits against the companies currently listed on GetDeal in the same sector. It benchmarks against this platform, not the whole market, and is withheld when the cohort is too small.
- get_industrial_scorepublic:read
How AI-dependent the companies listed in a sector declare themselves to be, as a 0-100 index with the distribution behind it. It measures GetDeal listings, not the industry, and it is not a forecast of automation potential.
Deal discovery
Live listings and the investor directory. Anonymization is applied per caller, on the server, to every one of these.
- search_dealspublic:read
Search companies listed on GetDeal that are raising capital or seeking acquisition, filtered by industry, country, funding stage or track. A company selling confidentially comes back as “Confidential Company” with its financials withheld, and the response counts how many results those are.
- get_dealdeals:read
Full detail for one listing, at whatever depth the person holding the key is entitled to see. A confidential listing returns without its name, website or financials until an NDA is signed on that deal, and says so in a notice field.
- list_new_dealspublic:read
Listings created recently, newest first. It answers “what is new since I last looked” without an agent re-running a full search every time.
- similar_dealsdeals:read
Listings resembling a given one, matched on industry and funding stage. The match uses only those coarse attributes — never identity or financials — so it cannot be used to work out who a confidential seller is.
- search_investorspublic:read
Search GetDeal's curated directory of active investors by thesis, stage, country or free text. It is a hand-maintained public directory, not a list of the platform's users.
- get_investor_profilepublic:read
The full directory entry for one investor, including what they say they look for and what they bring beyond money.
Diligence
The AI analysis GetDeal already holds for a listing, returned at the depth the caller is entitled to. These read a stored analysis; they do not commission a new one.
- analyze_startupdiligence:run
Success probability, overall score, valuation model output and market position for one listing. Before an NDA is signed you get the summary and the detailed figures are withheld rather than absent — a depth field says which of the two you received.
- risk_reportdiligence:run
The concerns the stored analysis raised about a listing. Mechanical rather than judgemental: the absence of a flag is not evidence that there is no risk.
- compare_dealsdeals:read
Two to ten listings side by side on their listed attributes and analysis scores. A listing you cannot see is reported as not comparable rather than dropped, so a comparison never quietly covers fewer companies than you asked for.
Matching and actions
Thesis matching in both directions, and the first of the two tools that reaches a real person on the other side.
- match_startupsdeals:read
Companies raising that match an investment thesis. Give at least one criterion; the ones you leave out are neutral rather than a penalty. Fundraising listings only — confidential M&A listings are excluded from this surface entirely.
- match_investorsdeals:read
Investors from the curated directory whose stated thesis matches a listing you own. It belongs to the listing owner: a caller who does not own the listing is refused, and a wider key does not change that.
- express_interestactions:writeasks a human first
Registers your interest in a listing. This reaches a person — the company is told who you are and a conversation opens — so the first call asks your human to confirm. Below verification Level 1 the platform records a pending intent instead and contacts nobody until verification completes.
Data room
Strictly per deal, and strictly metadata. Document contents are not available through an agent at all — see the limits below.
- list_data_roomdataroom:read
The documents on a deal that you can see: file names, folders, sizes, and whether each one is required. Access follows the deal rather than the key, so a caller who is not a party to it opens nothing.
- data_room_activitydataroom:read
Who opened which document on a deal, and when. Useful for judging whether interest is real rather than polite.
Portfolio and listings
The caller's own side of the marketplace — an investor's pipeline, a founder's listings, and what is missing before a raise.
- get_portfolioportfolio:read
Your own deal pipeline as an investor: how many deals, at which stages, and — on a paid plan — the sector split. A thinner answer on a free plan is a plan limit rather than an empty portfolio, and the tier field says which.
- get_listing_analyticsportfolio:read
How your own listings are performing: how many investors are in conversation, and at which stages.
- readiness_checkportfolio:read
What GetDeal can see is missing before you raise. Observable account and listing signals only — it has not read your deck, and it is not a diligence checklist.
Deal lifecycle
Where a deal of yours stands, and the one call that moves it. The second of the two tools that reaches a person.
- get_deal_pipelinedeals:read
Where one of your deals stands: stage, status, whether the NDA is signed, and what that gates. Worth calling before wondering why a listing's details are withheld.
- advance_deal_stageactions:writeasks a human first
Moves one of your deals to its next stage. The other party sees it immediately and the agent cannot undo it, so the first call asks your human to confirm.
And one diagnostic
- healthno scope needed
Confirms the connection works and reports which capabilities the key grants. It takes no arguments and needs no scope — a diagnostic behind a scope gate cannot diagnose a scope problem. Call it first when something is not working.
Human in the loop
Actions that reach a person
Two of the twenty-five tools do something a person on the other side will see, and neither can be called on the agent's own initiative.
The company is told who you are, and that is how a conversation starts.
The other party sees the move immediately, and the agent cannot undo it.
How the confirmation works
First call
The tool does not act. It answers with a request for confirmation carrying a sentence written for a person — which company, what will happen, and that nothing has been sent yet — plus an opaque piece of state for the client to hand back.
The client asks the human
This is the protocol's own multi-round-trip mechanism, not a GetDeal invention, so a conforming client renders that sentence natively — the person sees the question, not a JSON blob they are asked to approve.
The retry
The agent calls again with the same arguments and the state echoed back. That state is signed and bound to the acting user, the tool name and a digest of those exact arguments; it expires, and it is good for exactly one execution. Only then does the action run.
An agent cannot take these actions on its own initiative
There is no argument, no header and no scope that skips the question. A key carrying actions:write lets an agent ask; it does not let it act. And because the confirmation is bound to the arguments that were approved, an agent that edits the deal id between the two calls does not get a confirmed action pointed somewhere else — it gets asked again. One approval authorises one action, not every repetition of it.
The reason is concrete. An agent screening two hundred listings overnight and expressing interest in forty of them would put its human's name in front of forty founders without that human ever seeing the list.
Beyond tools
Reference material and named workflows
Three resources an agent can read, and five prompts it can run. Neither reads a specific company: everything that does goes through a tool, where the scope gate and the audit ledger are.
Resources
Reference data an agent reads rather than calls. The taxonomy exists because our names are literal — an agent guessing at one gets an empty result and reports it as an absence.
getdeal://reference/taxonomy
GetDeal taxonomy
The exact industry, sub-industry, funding stage and instrument names GetDeal uses. Filters match these strings literally, so a near miss like “FinTech” or “Series-A” returns nothing and reads to an agent as “there are no such companies”.
getdeal://reference/valuation-methodology
How GetDeal values a company
The multiples, the growth and margin adjustment factors, the date the comparables are as of, and what the method is not: no company-specific diligence, no cap table, no quality of earnings.
getdeal://reference/answer-format
How to read a GetDeal answer
What each envelope field means, the difference between a withheld field and an absent one, the full scope and error vocabularies, and the scopes the calling key actually carries.
Prompts
Ready-made workflows your client can offer by name. A model asked to invent the sequence for an investment memo invents a different one each time; a named prompt makes the good sequence the default.
- screen_deal_flow
Screens live deal flow against a thesis you describe: read the taxonomy first, search, pull detail on anything promising, then rank the results and say why each one places where it does.
- prepare_investment_memo
Assembles a first-draft memo for one listing — company, market, traction, comparables, indicative valuation, risks, open questions — with every figure carrying the source and as-of date the tool returned.
- fundraise_readiness
Places a company against its sector cohort: how active the sector is, where its ARR sits, an indicative range, and what an investor will ask for that is not there yet.
- compare_deals
Sets several listings side by side for an investment committee against criteria you state — one table, then a short recommendation, with confidential listings compared on what is visible rather than dropped.
- weekly_deal_digest
Summarises what has newly listed against your criteria, grouped by how well each one fits, short enough to read over coffee — and says so plainly when nothing fits.
Setup
Connect it
One endpoint, one header. The server speaks the current Model Context Protocol revision and still answers older clients on the same URL, so most MCP clients need nothing beyond the URL and a key.
- Endpoint
POST https://api.getdeal.ai/api/mcp- Authentication
Authorization: Bearer <key>. The endpoint is deliberately blind to browser session cookies — a bearer key is the only credential it accepts.- Keys
- Created in your GetDeal dashboard, with a label, a chosen set of scopes and an optional expiry. The secret is shown once, at creation, and is never retrievable afterwards — store it the way you store any other credential.
Claude Desktop, or any MCP client
Add one server entry to your client's configuration file and restart it.
{
"mcpServers": {
"getdeal": {
"type": "http",
"url": "https://api.getdeal.ai/api/mcp",
"headers": {
"Authorization": "Bearer gd_mcp_xxxxxxxxxxxx_..."
}
}
}
}Or straight from a shell
Listing the tools is the quickest way to confirm a key works and to see exactly what that key is allowed to call.
curl -sS https://api.getdeal.ai/api/mcp \
-H 'Authorization: Bearer gd_mcp_xxxxxxxxxxxx_...' \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
-H 'MCP-Protocol-Version: 2026-07-28' \
-H 'Mcp-Method: tools/list' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{"_meta":{
"io.modelcontextprotocol/protocolVersion":"2026-07-28",
"io.modelcontextprotocol/clientCapabilities":{}}}}'The full request and response shapes, the worked confirmation exchange, the error codes and the rate limits are in the developer quickstart in the repository at docs/mcp-quickstart.md.
Scopes
What a key is allowed to do
Six scopes, all six now in use. You pick them when you create the key, and a key carries only what you gave it — so the natural thing to do is issue a narrow key per agent rather than one key that can do everything.
- public:read
- Reference data and public listings — the freemium surface, teasered exactly as the public marketplace is.
- deals:read
- Deal and listing detail, at whatever depth the acting user already has.
- diligence:run
- The AI analysis pipeline. The expensive tier, and the one to grant most deliberately.
- dataroom:read
- Data-room documents, strictly per deal and never beyond the user's own grants. Metadata only; contents are never returned.
- portfolio:read
- The acting investor's own portfolio and its monitoring signals.
- actions:write
- Anything with a consequence — expressing interest, moving a deal stage. Every one of these is human-confirmed.
Provenance
What you get back
Every tool returns the same envelope around its answer. It is declared as each tool's output schema, so any conforming client can check it — this is a contract, not a convention.
- source
- Where the answer came from — a valuation model, a set of listings, a reference table.
- asOf
- The date the underlying data was true as of, so an agent can weigh how stale it is instead of assuming it is current.
- methodology
- How the answer was derived, in one sentence an agent can quote back to its human.
- confidence
- A number from 0 to 1, or null where confidence is not a meaningful idea for this answer. A valuation range scores low on purpose: it is arithmetic on one figure you supplied.
- withheld
- The fields GetDeal holds and this caller was not permitted to see. Empty when nothing was held back.
Read this one twice
Withheld is not the same as absent
A company selling confidentially on GetDeal has a name, a website and financial figures. They are withheld from you until an NDA is signed on that deal. They are not missing, and the company is not a company without financials.
Withheld
The field is named in withheld. GetDeal holds this value and you may not see it yet. Sign the NDA on that deal and the same call returns it.
Absent
The field is simply not in the response and not in withheld. GetDeal does not hold this value at all. No signature will produce it.
Collapse the two and every company your agent cannot fully see becomes, in its summary, a company with nothing to show — which is both wrong and unfair to the seller. So a pre-NDA sale card reports its company name, website, logo, amount sought and ARR in withheld, and get_deal adds a notice in plain language saying those values exist and what unlocks them. The AI analysis does the same: before an NDA it returns a summary and names the withheld figures rather than returning nulls that read as findings.
Limits
What this does not do
An agent repeats what it is given. So the limits travel in the payload as well as on this page.
A valuation range is not a valuation
It applies a sector revenue multiple to the one revenue figure you supply, scaled by a growth band and a margin band. No company-specific diligence, no cap table, no quality of earnings. It is not investment advice, and the midpoint of the range is not a price.
Small cohorts are suppressed, not estimated
Aggregates are computed over real listings, and an average over one listing is that listing wearing a statistic's clothes. Below five matching listings the tools report that the cohort was too small rather than returning a number — an agent can issue those queries far faster than a person, and differencing two of them must not isolate a company.
Document contents never reach an agent
The data-room tools return metadata — file names, folders, sizes, who opened what and when. They never return the document. Downloads are watermarked per viewer, and a watermarked file read into a model's context stops meaning what the watermark says, so the content is named as withheld and the person reads it in GetDeal.
An agent never sees more than its human
A scope narrows, it never widens. What a call may actually do is the intersection of the scopes granted to the key and what the account behind it already has — role, verification level, NDAs signed, per-deal grants. A key carrying dataroom:read, held by someone who has not signed the NDA on a deal, opens nothing on that deal. Handing an agent a key is strictly weaker than handing over a password.
The tool list itself depends on the key
Two agents connected to the same endpoint can be shown different tools, because the list is filtered by the scopes on the credential presented with the request. A key without portfolio:read is not offered get_portfolio at all, rather than being offered it and refused. So the answer to "which tools are there" is always: call tools/list with your key and see.
Benchmarks are against this platform
Market snapshots, benchmarks and the industrial score describe the companies listed on GetDeal, not the whole market. Listed figures are as stated by the company and are not independently verified. The envelope says which is which on every call.
Create a key and connect your agent
Keys are created in your GetDeal dashboard: give it a label, pick the scopes it should carry, and copy the secret once. It takes about a minute, and health will tell you immediately whether it worked — and which of the six scopes the key actually carries.
No account yet and just want to see the model? Try the free valuation or read the FAQ.