Docs

Amdahl documentation

Ask your customer conversations a question and get an answer in one call, hand a multi-step investigation to an agent, or grade what you are about to send against what your buyers actually said.

QuickstartAPI reference
app.amdahl.ai/api/platform/v1Two doors
POST /search/queryblocking

Ask, get rows back

Plain language, typed filters, or meaning. Returns the rows and the SQL it ran, in about a second.

POST /chatasync

Hand over an investigation

Returns handles immediately. A Master agent works your corpus and cites every figure it reports.

Amdahl turns your CRM, call recordings and meeting notes into normalized customer intelligence, then gives you a small set of primitives to build with.

What is ours, what is yours

Amdahl

  • Search your own conversation corpus - typed filters, plain language, or meaning
  • Grade a prompt or a message against real customer quotes
  • Run an Agent over all of it, on demand or on a cron

You

  • the prompts and the canon behind them
  • the orchestration around the calls
  • the approval flow before anything ships

There are two decisions ahead of you, and they are independent. First how you connect - REST or MCP, the same operations either way. Then what you call - a synchronous endpoint or an agent.

How you connect

Build with the API

Authenticate a script or a backend service with an API key and call Search, Eval, Chat and Routines over plain REST. Start with the Quickstart.

Connect an AI agent

Point Claude, Cursor or ChatGPT at Amdahl's MCP server and let an agent work your customer corpus directly - approved in your browser, no key to copy.

What you call

Endpoints

Synchronous verbs, one blocking call each. Search your warehouse (plain-language, typed filters, or meaning), and Eval a draft against your own customer quotes.

Agents

Anything that needs a multi-step, server-side investigation. The Chat ask-door runs a Master agent over your workspace data, the agent library stores the agents you build, and Routines run them on a cron.

Reach for an endpoint when you already know the shape of the answer: a filtered slice, a meaning-based ranking, a grade on a draft. Reach for Chat when you want Amdahl to work out the shape for you.

The endpoint surface

POST/search/queryauth required

One routed door over your conversation corpus - typed filters, plain-language NL to SQL, or meaning-based semantic matching. Returns the rows and the SQL it ran.

POST/evals/runauth required

Grade a prompt and a drafted message against verbatim quotes retrieved from your own calls and emails. Async - returns a handle you poll.

POST/chatauth required

Hand over an investigation. Returns run handles immediately; stream or poll for the answer.

How the verbs chain

Each endpoint is one blocking call, which is what makes them composable. The canonical chain is find the cohort, then read what it said:

Two calls, no agent run

  1. Search

    "Closed-won deals this year, largest first" - real pipeline company ids.

    mode=filter

  2. Search again

    The objections those accounts already raised, in their own words.

    mode=semantic

Both lanes go through one door, so the second call is the first one with a different mode. When the ask needs judgment rather than retrieval, or outside market signal, hand it to Chat.

Guides

The grading loop

Write a prompt, ground it in real customer evidence, draft, grade, fix, rerun. The method, what to measure, and the two caches that will otherwise hand you a replay and let you call it a measurement.

Organizing your bench

A run is the unit and finished runs are immutable. The folder layout, the manifest that makes a run reconstructable, and the disagreement log that tells three different failures apart.

Reference

  • API reference - the live REST reference generated from our OpenAPI spec, with a built-in "try it" console.
  • Tool catalog - every operation an agent sees, projected from the operation registry.
  • Errors - the shared error envelope and the gate codes worth handling.