Docs

Onboard with an agent

Paste one short prompt into the agent you already use (Claude, Cursor, ChatGPT, Codex or Claude Code). It finds the people you are about to write to, gathers what it knows about each one, has Amdahl write a few messages and shows you each one with every try, then, once your conversations are synced, runs a search for real customer quotes and evaluates a draft against them.

The fastest first result is to let the agent you already use do it. It already knows your tools: the sequencer or SDR tool you connected to it, your CRM, your Drive or Notion, your repo. Paste one prompt, and it finds the people you are about to write to there, gathers what it knows about each one, and has Amdahl write a few messages. Amdahl writes from that context, and a draft is optional, so the agent's job is to gather the context well. See Context. The steps live in the onboarding skill, which the agent loads and follows.

The agent changes nothing in your tools on its own. It shows you the messages first and only writes them back to your sequencer, CRM, files or code after you approve that batch. It never sends anything.

Only want your emails written?

Paste one line instead: Load and follow https://docs.amdahl.ai/skills/amdahl-optimizer/SKILL.md. The amdahl-optimizer skill is the shorter first run: it finds a few people you are about to email, has Amdahl write to each and shows every try, with no data connected. This page's prompt does that too, then goes on to Search and an evaluated draft once your conversations are synced.

Have a whole campaign?

An agent in a chat is for a few messages: three to five. For a table of 50 or 500, use the campaign cookbook: a short script that builds the context for each row, runs the whole list with an API key, and can stop and resume. The agent tells you this when you point it at a long list, and offers a sample first.

Before you paste

  • A workspace. Join your team's workspace, or ask an admin to add you. No workspace yet? Try Amdahl at console.amdahl.ai/try, which also puts you on the beta waitlist.
  • Amdahl connected to your client. For Claude Code, Codex or Cursor, run npx amdahl, sign in, and do the last step it prints. For Claude on the web or desktop, ChatGPT and others, see Connect your agent. There is no key to copy. The first time the client calls an Amdahl tool, it may ask to allow it: choose Allow once or Always allow.
  • Your recipients, and what you know about them, wherever they are. A CRM, a lead list, a sequencer. Nothing to export. If the agent can't find them, it asks you. It also needs to know what you sell: your site or a one-pager is enough.
  • For the evaluated draft, synced conversations. The agent checks and only offers it when there is data to evaluate against. On a workspace with no connected source, an evaluation's four evidence criteria abstain and only the ask is evaluated. Amdahl still writes messages there, evaluated without your data. There is no shared demo workspace today.

To check the setup at any time, ask the agent to call get_setup_status. It names the workspace you connected and says whether optimizing will work, and if not, why. See Check your setup.

The prompt

For Claude (web, desktop or Code), Cursor, ChatGPT or any client where the Amdahl MCP server is connected:

text
I've connected the Amdahl MCP server. Load and follow https://docs.amdahl.ai/skills/amdahl-onboarding/SKILL.md
I want Amdahl to write my outbound emails. Use what you already have access to and find a few people I am about to email, and what we know about each one. Ask me if you can't tell.
Then have Amdahl write to each of them from that context. Show me what Amdahl chose first, exactly as it came back, then the facts it was written from and each try in between.
Don't change anything in my tools or send anything without my OK.

That is the whole prompt. Once your conversations are synced, the agent also offers a search and an evaluated draft, and shows each criterion with its verdict and reason.

Where your recipients live

The agent looks at what it is already connected to before it asks you anything. What it does in each common setup:

Your recipients are inWhat the agent does
A sequencer or SDR tool with an MCP serverReads the next few people due a step in the sequence or campaign you name, gathers what the tool knows about each, has Amdahl write each message, and shows it next to the step it came from. Writes approved messages back to the same step only when you say so.
HubSpot or SalesforceReads the contact, the company, the deal and the notes through the CRM connection it has, and sends them as facts. Writes nothing back to the CRM without your approval.
A spreadsheet or Excel fileReads the newest rows, turns each row's columns into facts, and offers the messages as a new column next to your rows.
A Google Drive or Notion folderFinds what you sell and the proof you may cite, uses your style guide as rules and emails you liked as voice_examples, so each message keeps your voice.
Code (Codex or Claude Code with your repo open)Finds where your code knows the recipient, writes a few real messages, then proposes a diff that calls POST /api/platform/v1/messages/optimize (or the MCP optimize_message tool) with that context before each email is queued. It does not commit or push without your OK.
Nowhere in particularAsks you to name one person and why you are writing, and starts there.

It asks at most three short questions, only for what it couldn't find: who to write to, why now, and what you offer.

The prompt: with an API key

For a headless agent, or a coding agent without the MCP server connected. Create a key with amdahl keys create --name "Onboarding" --preset agent (see Command line (CLI)) and put it in the agent's environment as AMDAHL_KEY. A coding agent (Claude Code, Codex, Cursor) or a script reads the key from there, so do not paste it into its chat.

In a private chat with an AI assistant such as Claude or ChatGPT, pasting the key is fine. The console's Copy key + prompt button does this, because those apps have no other way to receive a key. The key reaches one workspace only, and you can revoke it at any time. If you share or export that chat, revoke the key in the console under Settings, then API keys.

text
Set me up on Amdahl and run my first use case end to end.
My Amdahl API key is in the AMDAHL_KEY environment variable.
Use case: grounded outbound. I sell to {persona, e.g. "VP of RevOps at mid-market SaaS"}.
Load and follow the onboarding skill at https://docs.amdahl.ai/skills/amdahl-onboarding/SKILL.md
First win: find a few people I am about to email with what you can access (or use this one), and have Amdahl write to them:
"""
{who the recipient is, why you write now, and what you want them to do, or leave this out}
"""
Then check my data; if conversations are synced, offer one search and one evaluated draft, and show every criterion with its verdict and reason.
Explain as you go; do not build anything beyond these calls.

Leave the recipient out and the agent looks for people to write to, then asks you to name one if it finds none. Leave the use case out and, once your data is synced, it offers you the two to choose from.

If your outbound runs in code

Open the repo that sends your emails. The agent writes a few real messages first, then shows you where the optimize call goes and the diff it would make. The call sends what your code knows about the recipient as context, not a draft. Each call usually takes 1 to 2 minutes and can take up to 300 seconds, so the call belongs in a background job before the send, with a fallback of your own when ok is false. Optimize before you send has the code. You review and commit the change yourself.

What the agent does

Connect

Uses the MCP tools if they are there, otherwise REST with your key.

Check your data (optional)

Reads GET /search/overview: rows per surface and the newest conversation date. This only decides what comes after the first win; an empty workspace does not stop it.

Find your recipients

Checks the tools it is connected to (a sequencer, a CRM, Drive or Notion, a spreadsheet, a repo) for people you are about to write to, and what is known about each one. Asks you only what it can't find. Reads only.

Write a few

The first win, no data needed. Sends 10 recipients to POST /messages/optimize with their facts in context and "include_tries": true, one call each (usually 1 to 2 minutes apiece). For each recipient it shows Amdahl's choice first, exactly as returned, with its one-line summary, then the facts it was written from and every try Amdahl scored on the way (its score and Human tone, what it aimed at, and which one came back). Those scores are Amdahl's own scoring during the search: they show the path, not proof (see Reading tries). When the context was too thin, it puts Amdahl's questions to you. It shows the returned message exactly as it came back and never rewrites it with its own model; to change it, it calls Amdahl again with your wish as a rule or a fact. It writes approved messages back only after you say yes to that batch.

Pick a use case (once synced)

Only when the workspace has synced conversations. Grounded outbound (draft a first touch in your buyers' words) or a message evaluation (evaluate an email you already wrote against your customers' own conversations). On an empty workspace it stops here, tells you to connect a CRM or call recorder, and says to come back after the first sync, rather than returning empty results that look like answers.

Search

One semantic search for 5 to 10 verbatim customer quotes on the topic, with the company and date for each.

Evaluate

Evaluates the message with optimize_message in mode: "grade" with "evidence": "workspace", which returns on the same call. It shows how many criteria the message passed (for example 4 of 5), then each criterion with its verdict and reason. See What an evaluation measures.

Close the loop

Asks what you did with the message and records it with record_message_feedback, then offers the next step: evaluate the prompt behind your drafts with optimize_prompt, or put an evaluation in front of your send queue, so each message is evaluated before it ships.

If something goes wrong

You seeIt means
401The key is missing, wrong, revoked or expired. Create a new one.
403 forbidden, "Missing required scope(s): messages:execute" (over MCP: "optimize_message requires the messages:execute scope")The key is Read only or older than the optimize call. Create a Customer agent key. Over MCP, disconnect and reconnect Amdahl.
429 quota_exceeded on the optimize callThe workspace used this month's 1,000 optimize calls for API and OAuth keys. It resets next month.
The message is genericThe context was thin. Add a reason to write now and what you offer as facts, then run it again. See Context.
The result has rows in omittedThe context could not back part of what the message should say, so the message leaves it out. Each why says what is missing. Add it as a fact, then run it again.
outcome: "needs_input"Only with on_missing_input: "ask": the context could not support the request, and nothing was written. Answer the questions in ask, then run it again.
"Your workspace has no synced conversations yet"Amdahl still writes messages. For the search and the evaluated draft, connect a CRM or call recorder in the console, wait for the first sync, then run the prompt again.
A criterion has pass: nullThat criterion was not evaluated. Its not_applicable_reason (for example corpus_empty) says why; it is not a failing criterion. When no criterion was evaluated, score is null.

Doing it yourself instead

The Quickstart makes the same calls with curl, one at a time. Which call when maps a job to the call that serves it.