AI and outreach

Chat agent

The chat agent is a Claude-powered assistant built into the dashboard. You describe what you want in plain language, and it does the work with the same tools you would use by hand: searching leads, updating records, checking jobs, reading analytics and managing sources.

What it is for

Use the agent when a task is quicker to describe than to click through. It works on your real data, not on a copy, so anything it changes shows up in the rest of the dashboard straight away. It does not invent results: when it needs a fact, it fetches it with a tool.

Things you can ask for:

  • "Find all leads from Acme Corp with valid emails."
  • "Which of my sources produced the most leads this month?"
  • "Why did last night's scraping job fail?"
  • "Mark the leads tagged conference as contacted."
  • "Create a campaign from this brief" with a PDF attached.

Have your first conversation

  1. Open AI Assistant in the dashboard sidebar.
  2. Type a request and send it. The reply streams in as it is written.
  3. Watch the tool calls. Each time the agent uses a tool, the conversation shows which one it ran and what came back, so you can check its work.
  4. Follow up in the same conversation. The agent keeps the earlier messages as context.

You can attach a PDF or a CSV to a message, up to about 5 MB. A message can be up to 10,000 characters. If you write in a language other than English, the agent replies in that language.

How it runs tools

The agent has tools for leads, sources, jobs, campaigns, profiles and matching, analytics, enrichment, warmup domains, webhooks, workflows and team collaboration. For each request it picks the tools it needs, runs them in order, and writes its answer from the results. One request can involve several tool calls.

Approvals

The agent is built to pause before operations that are destructive, that consume credits, or that touch many records at once. When a request needs sign-off, an approval appears in a banner above the conversation with approve and reject buttons. Nothing runs until you approve it.

In a team, a member who is not an admin can ask the agent for a bulk delete, export or update; the agent files an approval request and the team admin is notified to review it.

Conversations

Every chat is saved as a conversation that you can return to, search, and delete. A few things that help once you have more than a handful:

  • Templates start a conversation that is already set up for a common task.
  • Branching forks a conversation so you can try a different direction without losing the original.
  • Sharing and export let you hand a conversation to a teammate or keep a copy.
  • Memories are facts the agent picks up from your conversations, such as a preference for a region or a segment, and recalls later when they are relevant.

Use it from the API

The agent is also available over the API. Create a conversation, then post messages to it. The reply is a Server-Sent Events stream.

# 1. Start a conversation
curl -X POST https://scrappy.gg/api/chat/conversations \
  -H "Authorization: Bearer sk_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"title": "Lead research session"}'


# 2. Send a message and read the reply as a stream
curl -N -X POST https://scrappy.gg/api/chat/conversations/CONVERSATION_ID/messages \
  -H "Authorization: Bearer sk_your_api_key" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"content": "Find all leads from Acme Corp with valid emails"}'

The stream carries one event per step:

event: text
data: {"delta": "I'll search for leads from Acme Corp..."}


event: tool_call
data: {"toolName": "search_leads", "input": {"company": "Acme Corp"}}


event: tool_result
data: {"toolName": "search_leads", "result": {}}


event: done
data: {"messageId": "msg_123", "usage": {"inputTokens": 500, "outputTokens": 150}}
EventMeaning
textA piece of the reply. Append each delta to build the full message.
tool_startThe agent has begun running a tool.
tool_callThe tool name and the input the agent gave it.
tool_resultWhat the tool returned.
doneThe reply is complete. Includes the message id and token usage.
errorThe request failed. The data carries the reason.

Pending approvals are listed at GET /api/chat/approvals and resolved with PATCH /api/chat/approvals/{id} and a body of {"approved": true}. Chat messages are limited to 1,000 per hour per user. If you use your own AI client, the MCP server exposes the same platform as tools.

What it costs

Chat is metered at 1 credit per 1,000 tokens, counting both what the agent reads and what it writes. A short question costs a fraction of a long research session. Tools the agent runs on your behalf cost what they normally cost: a scraping job is still 10 credits and an email validation is still 1. See Credits and billing for the full list.

Where to go next