Agents

Agents let AI operate Starlight for you. Open the panel with the Agents button (or press A) on any page to chat with an assistant that can search leads, judge their fit, draft outreach, and — with your sign-off — send it. You can also create virtual SDRs: persistent agents that run on a schedule and work even while you sleep.

Chat

The Chat tab is a command surface for everything in Starlight. Ask things like "find 10 leads that match our ICP and judge their fit" or "draft personalized outreach to my best unsent contacts." The agent uses the same tools you do — every search lands on the Search page, every draft waits in Messages.

Type @ in the chat box to tag a specific contact or company. A search opens over your Contacts and Companies; picking one adds a chip to the composer so the agent knows exactly which record you mean — "make the @Ron Gabrisko messages more concise" goes straight to that person's drafts, even if two contacts share a name.

Before recommending any send, the agent runs an independent quality judge over each draft. It will always ask for your explicit confirmation in the conversation before approving anything for sending.

You can stop the agent at any time. While it is working, the send button becomes a Stop button (or press Esc). Stopping halts the agent immediately: no further steps run and no further blocks are spent. A step that was already underway is allowed to finish rather than be cut off halfway, so the conversation tells you exactly where things landed — what completed, what was interrupted, and that nothing new was started.

The chat agent also knows your workspace setup — whether your email and calendar are connected, whether the Chrome extension is installed and active, and whether your booking page is ready — so it can flag gaps before a campaign instead of guessing. And it can manage your virtual SDRs directly from the conversation: ask it to create an agent, adjust a quota or schedule, pause or resume one, kick off a run right now, or report what an agent did on its last run. A clear ask is all it needs — request an Autopilot agent and it creates it that way in one step, telling you plainly that passing drafts will send without review. It only asks follow-up questions when your request leaves out something material, and deleting an agent always requires your confirmation.

It can also work your calendar directly: ask “what times am I free Thursday?” and it checks your real availability (schedule, connected calendar, buffers, and booking caps), then books the meeting for you — creating the calendar event with a meeting link and sending confirmation emails to both sides, with the meeting linked to the contact automatically. It always confirms the guest and time with you before booking.

AI setup assistant

New to Starlight? The first item in the sidebar's Getting Started checklist, Let AI set up your outbound, opens the chat and asks the agent to build your whole outbound system for you. You can also just ask in chat any time ("set up my outbound system").

The agent looks at your current setup first: if you have nothing, it builds everything; if you already have a setup, it optimizes what is there. It presents a short plan with an estimated block cost, and nothing runs until you approve it. The centerpiece is your Strategy: one well-organized document laying out how to win your ideal customers — the objective, the audiences worth pursuing, the angle for each, weekly volume targets, and what to watch. The run then creates the recurring agents that execute it, with Autopilot on by default so quality-judged messages send automatically once your email is connected. It also fills gaps in your company profile, runs a first lead search, judges fit, adds the best matches to Contacts, drafts a first outreach batch (those land in Messages as pending review), and sets up signal tracking. A hard block budget caps the spend, and steps that need your browser, like connecting email or calendar, are listed for you at the end.

The strategy lives under Settings → AI as a single rich-text document you can edit directly — or just tell the chat what to change. Every Monday an AI strategy review reads your results: leads found, fit rates, draft quality, replies, and deliverability. It tunes your agents toward what works, rewrites the strategy document so it stays accurate, and leaves a plain-language summary of what changed (2 blocks per review; it never starts paid work on its own and never touches your Autopilot settings).

Virtual SDRs

The My Agents tab manages persistent agents. Each one has instructions — one freeform document where you write exactly what you want it to do: the goal, who to go after, and how to pitch — plus a weekly lead quota and a schedule. On each run the agent:

1. Sources leads from its query set — paging deeper into searches that are already working, and planning fresh complementary angles only when those are drained

2. Scores every result 0–100 against your company's ideal customer profile and the agent's instructions, keeping only strong fits (50+ by default) and skipping anyone it has already contacted

Agents never write to people who work at your competitors. Anyone employed by a company on your competitor list — the Competitive landscape section of your product profile under Settings → AI — is dropped before drafting, and never charged for. The same exclusion covers signal and social auto-drafts and autopilot approvals. Drafting to one on purpose still works from chat or the Outbound builder, with a warning.

3. Adds the keepers to Contacts and drafts a personalized sequence for each

4. Scores every draft for personalization grounding, tone, and spam signals

Because the agent paginates a query that fits instead of reinventing one each run, well-targeted instructions keep producing fresh, relevant drafts run after run rather than running dry after the first page. It only pauses itself once it has genuinely exhausted every angle — if that happens, broaden the instructions or lower the fit bar and resume.

Set the weekly lead quota to match the volume you want; higher quotas contact more new people each week. Sustained high volume also depends on your sending capacity — see Deliverability.

Every run produces a timeline you can audit — the searches it ran, the judge scores, and why leads or drafts were rejected.

Autopilot

By default, everything an agent drafts waits in Messagesfor your review — nothing sends without you. Flip an agent's Autopilot toggle and it will approve and send drafts that pass its quality bar on its own:

• Email sends from your connected Gmail or Outlook account on the normal schedule
• LinkedIn requires the Chrome extension with its own Autopilot toggle enabled — while Chrome is open, the extension sends approved LinkedIn messages automatically, capped at 25 per day to protect your account

Use Starlight from Claude Code, Codex, and other AI tools

Starlight ships a hosted MCP server, so any MCP-capable AI tool can run your prospecting and outreach. In Claude Code:

claude mcp add --transport http starlight https://joinstarlight.com/api/mcp

Your browser opens, you sign in to Starlight (or create an account), and approve the connection. Then just ask your AI tool things like "find 20 leads for my company" or "draft personalized outbound to 20 leads." Drafts are created for review; sending always requires the assistant to confirm with you first.

Building your own integration? The whole surface is described at /openapi.json (OpenAPI 3.1: the OAuth endpoints, the MCP endpoint, and every tool with its input schema) and summarized in the MCP manifest at /.well-known/mcp.json. OAuth clients can ask for starlight:read (free, read-only tools), starlight:write (everything else), or the default starlight:agent (both). Every API error is JSON with a stable code and a hint.

Browser agents get the same tools with zero setup. When you're signed in to the dashboard, Starlight registers its tools with the emerging WebMCP standard (document.modelContext), so an AI assistant built into your browser can search leads, draft messages, and check your inbox through real tools instead of clicking around the screen. This is early: it currently requires a browser with WebMCP support (Chrome and Edge are trialing it), and it activates automatically where available. The same rules apply as everywhere else in Starlight: drafts are created for your review, and nothing sends without approval.

Blocks

Agents use the same block balance as the rest of Starlight. A scheduled agent that can't afford its next run skips it and retries on the next schedule.