Leadsmith

Find buyers, write outreach they reply to, and send it from your Gmail,
from a local dashboard powered by the coding agent you already use.

Get started · How it works · Tour · Safety · Compliance · Report a bug

No API keys. No paid data providers. No SaaS subscription. Everything runs on your computer, on your own Claude, ChatGPT or Google plan.
_{One optional paid add-on, BetterContact email verification, stays off unless you add a key.}

_{Screenshots use a demo workspace. Every company and person in them is fictional.}

What makes it different

Most cold email pitches. Leadsmith offers. For every lead, the agents find a real, dated reason to reach out now, then write three ideas made for that company. The prospect answers with a single letter:

Hi Daniel,

Stackwise launched v2 on Product Hunt (#2 of the day) with screenshots only and no demo video.

Three things I could make for Stackwise:
A) 45-second v2 walkthrough: replace the screenshots on the launch page
B) Three vertical launch clips: one feature each, for X and LinkedIn
C) Changelog video series: a 30s clip for every release

Which one should I make first: A, B or C?

"B" is the easiest reply anyone can send, and it tells you exactly what to build.

Prefer a classic, personal note? Switch Business profile → Email steps → Email style to Personal pitch + deck. The writer opens with one specific, true observation, says what you do for buyers like theirs, builds a 4-page PDF deck made for that company and attaches it, then asks for a short call:

Subject: Kestrel trail shoe, before the spring drop

Hi Dana,

The new Kestrel trail shoe looks great, and the weeks before a drop are when product video does the most work.

Northlight Studio films footwear and outdoor brands on real trails, so buyers can see grip and fit before they order.

I've attached a 4-page deck made for Kestrel with a two-day shoot plan.

Would a short call this week help?

How it works

• The engine is the coding agent you already use: Claude Code, Codex, OpenCode or Gemini CLI, running headless on your own plan. Connect one in Settings → Agent engine. Dashboard buttons start a background run, which you can watch live under Agent runs. If an agent hits its usage limit, finish the task with another one in a click.
• Four agents share one memory: a local SQLite pipeline plus your business profile.
  1. Signal Scout finds companies that fit your ICP and show a reason to buy now: hiring, funding, new leaders, launches, tech.
  2. People Finder finds the decision maker and their email. It looks for an address published on their site first and only guesses from the company's pattern as a fallback.
  3. Account Researcher writes a short, sourced brief and the outreach angle.
  4. Outreach Writer writes a 3-step sequence in your voice and your chosen style (an idea menu, or a personal pitch with a deck made for them), plus an optional LinkedIn note.
• You stay in control. Agents can research and draft, but they can never approve or send. Approving and sending are your clicks in the dashboard.

Free data only

Source	Gives you
Company websites	what they sell, pricing, team page, tech stack (60+ tools), socials, public emails, content freshness
Greenhouse / Lever / Ashby job boards	every open role, grouped by function: the strongest "need it now" signal
Google News	funding rounds, launches, leadership changes, expansions
Y Combinator directory	~5,000 startups by industry, region, team size and hiring status
Hacker News "Who is hiring"	hundreds of hiring companies every month, with role and website
Your agent's web search / fetch	everything else: lists, directories, press, podcasts
DNS (MX) + email patterns	a likely email per person, labelled guessed until it's proven
Your CSV	import leads from Apollo, Sales Navigator, Clay, a CRM or a spreadsheet

Optional add-on	What it does	Cost
BetterContact	A Verify button on guessed emails: checks many providers and returns a verified address	Your BetterContact credits, charged only when it finds a usable email. Off unless BETTERCONTACT_API_KEY is set. One person per click; agents never spend credits.

A tour of the dashboard

Leads: who to contact and why now

Every company gets a Fit score (how well it matches your ideal customer) and an Intent score (how strong and recent its buying signals are), with the one-line reason to reach out now. Flags such as EU, UK and CA mark leads in regions with stricter email rules. A ✓ means you marked the lead a good fit.

Use Find buyers to describe who you want in plain English, Add companies to paste domains, or Import CSV to bring your own list. Sources shows which discovery plays produced the leads you liked and the replies you got.

Lead details: research, signals and ideas

Click a company to see the research brief, every signal with its source link, the people found, and the three ideas offered. Mark a lead Good fit or Not a fit, with a reason if you like. The scout reads these verdicts before every new search and looks for more like your good ones.

Review: nothing goes out until you approve it

Every draft waits here. You see why now, the ideas, all three emails (day 0, 3 and 7), where the email address came from, and region notes. Edit anything, Send test to me, then Approve. Drafts with red issues from the linter (spammy words, AI-sounding phrases, links in a first email, unfilled {placeholders}) can't be approved.

• Edit next to the address fixes a wrong or guessed email. It's then marked Entered by you.
• Attach deck and View deck show and attach the PDF deck made for that company.

Outreach: sending, follow-ups and replies

Approved emails go out from your own Gmail, only in business hours in each prospect's time zone, one a minute, within your warm-up limit. Follow-ups reply in the same thread when they're due, and stop by themselves when someone replies. Replies show up inline: "A" marks idea A as picked, and Draft reply writes your answer for review.

Changed your mind about an approved email that hasn't gone out? Regenerate rewrites that person's unsent emails, with an optional note such as "shorter" or "warmer". They come off the send queue straight away and wait in Review until you approve them again. Sent emails are never touched.

Business profile: what every agent reads first

Your company and offer, ideal customer, personas, and voice and rules. Draft from website fills a first version from your site. Agents quote only the proof written here and never invent their own.

Email steps: style, and AI-written or your own template

Pick the Email style for every first email: an Idea menu (three ideas they choose from) or a Personal pitch + deck. Then, for each step, either let the AI write a fresh email, with an extra instruction if you like, or use your own template. Templates fill in lead fields such as {first_name}, {company} and {ideas}. Any other {name}, like {personal_line}, becomes one sourced sentence the AI writes. A live preview shows the result on a real lead.

Agent runs and settings

Every run shows its status, live log, cost and which agent ran it. If a run fails or you stop it, Run again retries it, and Run with another agent finishes it on a different one, picking up from what's already saved. That's handy when one plan hits its usage limit. Failures are explained in plain words, such as "the model's servers are busy" or "you've hit a usage limit".

Settings holds your agents (each card shows the plan it signs in with), the model, your Gmail connection (with Test connection), sending hours, warm-up, the daily limit, signature and postal address, and the optional BetterContact key.

Agent runs	Settings

Install

You need one coding agent, installed and signed in (see Choose your agent engine), and uv. uv installs Python for you. Chrome or Edge is optional, for PDF decks.

# Windows (PowerShell):  powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
curl -LsSf https://astral.sh/uv/install.sh | sh

git clone https://github.com/HariNayan/Leadsmith.git

cd Leadsmith

uv sync

Open the dashboard. Run this from inside the Leadsmith folder, and keep the terminal open while you use it:

uv run leadsmith

From any other folder, point uv at the project instead. This works in PowerShell, cmd and bash:

uv run --directory C:/path/to/Leadsmith leadsmith

On Windows you can double-click Start Leadsmith.cmd in the project folder instead.

It opens http://127.0.0.1:4600. Closing the terminal stops the dashboard. Check your setup any time with:

uv run leadsmith doctor

Choose your agent engine

Leadsmith works with any of these. Install one and sign in once in a terminal. Then click Connect next to it in Settings → Agent engine. If you start a run with no agent connected, Leadsmith asks you to connect one first. Only one agent is connected at a time, and only after you connect it: Leadsmith never starts or checks an agent in the background, and after you restart the dashboard you connect again.

• Your plan, on each card. Settings shows what each agent signs in with, such as Claude Pro, ChatGPT Go, your OpenCode providers or a Gemini API key. It reads this from the agent's own files on your computer; nothing is started or sent.
• New or previous session. Connecting asks whether to start a new session or continue the previous one, which shows when it was last used and what it was doing. While an agent stays connected, its runs share one conversation, so it remembers what it did in earlier runs.

Engine	Install	Sign in	Notes
Claude Code	see the site	claude, then /login	Runs the four agents as real sub-agents. Shows the cost of each run.
Codex	npm install -g @openai/codex	codex login	Uses your ChatGPT plan. Live web search on.
OpenCode	npm install -g opencode-ai	opencode auth login	Any provider you've added (Anthropic, OpenAI, GitHub Models, local…). Shows cost when the provider reports it.
Gemini CLI	npm install -g @google/gemini-cli	gemini, then sign in with Google (or set GEMINI_API_KEY)	Uses your Google account.

Every engine gets the same Leadsmith tools and the same limits. It can use the Leadsmith tools and the web, but not shell commands, file edits, or the tools that approve and send email. Engines without sub-agents get the four agents' playbooks and do each role in turn. Model in Settings is a dropdown of the connected engine's models: Claude Code's Sonnet, Opus, Haiku and Fable; the models in your Codex account; every model from your OpenCode providers, grouped by provider; or Gemini's Auto, Pro, Flash and Flash-Lite. Default lets the engine choose, and Other… takes any model name.

Also chat with Leadsmith inside your agent (optional)

The dashboard is the main way to work, but you can also give your agent Leadsmith's tools and chat with it. Replace C:/path/to/Leadsmith with your folder.

Claude Code: the repo is a Claude Code plugin, with the agents and skills included:

/plugin marketplace add HariNayan/Leadsmith
/plugin install leadsmith@leadsmith

Or, from a clone: claude --plugin-dir ./Leadsmith. Then try /leadsmith:setup, /leadsmith:find-buyers 10 fintechs in India hiring sales or /leadsmith:dashboard.

Codex:

codex mcp add leadsmith -- uv run --directory C:/path/to/Leadsmith leadsmith serve

OpenCode: add this to opencode.json (in your project, or ~/.config/opencode/opencode.json):

{
  "mcp": {
    "leadsmith": { "type": "local", "command": ["uv", "run", "--directory", "C:/path/to/Leadsmith", "leadsmith", "serve"] }
  }
}

Gemini CLI: add this to ~/.gemini/settings.json:

{
  "mcpServers": {
    "leadsmith": { "command": "uv", "args": ["run", "--directory", "C:/path/to/Leadsmith", "leadsmith", "serve"] }
  }
}

Then ask, for example, "Use the leadsmith tools: read the business profile, then find 5 companies that just raised a seed round and save them."

First run (10 minutes)

1. Business profile: enter your website and click Draft from website, then replace the remaining TODOs with real details (proof, buyer signals, voice).
2. Settings: connect Gmail. Any personal @gmail.com works. Turn on 2-Step Verification, create an App Password at https://myaccount.google.com/apppasswords, paste it, then click Test connection. Add your sender name, signature and postal address.
3. Leads → Find buyers: e.g. "10 seed to Series A fintechs in India hiring their first sales rep".
4. Review: edit, Send test to me, then Approve.
5. Outreach: turn Sending on in Settings, then click Review & send. Follow-ups go out in the same thread when they're due and stop automatically when someone replies.

Tip: set Test mode in Settings to your own address while you try it out. Every email then goes to you, marked [TEST].

Safety rails

• Nothing sends unless you approved the draft, Sending is on, and you confirmed the preview.
• Drafts with red lint issues can't be approved.
• Agents may only cite facts with a saved source URL, or your own business profile.
• Emails go out only inside business hours in the prospect's time zone (default 9–17, weekdays), one a minute, with follow-ups counted in business days. Opening the app after a break never fires a burst of late follow-ups.
• Warm-up: a new sender starts at 5 emails a day and adds 3 a day up to your daily limit (default 25).
• Every email gets an opt-out line, a List-Unsubscribe header and your postal address. A do-not-contact list is checked before every send, and opt-out replies and bounces are added to it automatically.
• An email that says something is attached can't be approved or sent unless the deck PDF really is attached.
• Regenerate and changing a person's email both move their approved, unsent emails back to Review, so nothing goes out that you haven't seen in its final form.
• Guessed emails are clearly labelled. Send small batches first, since bounces hurt your sender reputation.
• The dashboard only listens on 127.0.0.1 and rejects cross-site and DNS-rebinding requests.
• Leads that look like they're in the EU/EEA, UK, Switzerland or Canada get a flag with that region's rule in one line.
• You are responsible for following email and privacy law where you and your recipients are. Read COMPLIANCE.md (a summary, not legal advice).

Where your data lives

~/.leadsmith/     (Windows: C:\Users\<you>\.leadsmith)
├── .env          Gmail + sending settings (edited from the dashboard)
├── brain/        company.md · icp.md · personas.md · voice.md
├── leadsmith.db  accounts, signals, contacts, drafts, ideas, runs (SQLite)
├── decks/        generated PDF decks
└── exports/      CSV exports (Instantly, Smartlead, Lemlist, HubSpot…)

Nothing leaves your computer except the emails you approve and the web lookups the agents make.

Exports use the column names Instantly and Smartlead map automatically (email, first_name, last_name, company, title, website, linkedin_url). After those come reason (why now), idea_a/b/c, step1_subject, step1_body, step2_body and step3_body, plus timezone and strict_region, all usable as custom variables. Opted-out and bounced people are left out.

Set LEADSMITH_HOME to keep a separate workspace per client or product. Coming from GTM Brain (the old name)? Your ~/.gtm-brain folder is moved to ~/.leadsmith on first start, with nothing lost.

Commands

Command	What it does
uv run leadsmith	open the dashboard (--port 4700, --no-open)
uv run leadsmith doctor	check the agent engines, free sources, PDF renderer, Gmail, profile
uv run leadsmith serve	MCP server over stdio (what your agent launches)
uv run leadsmith init	create the workspace without opening the dashboard

Development

uv run python -m unittest discover -s tests -v

claude plugin validate .

Layout:

src/leadsmith/
├── server.py        MCP tools the agents use
├── dashboard.py     local API · web/ is the UI
├── jobs.py          runs the chosen agent in the background
├── engines.py       Claude Code, Codex, OpenCode and Gemini CLI adapters
├── providers/       free data sources (+ optional BetterContact)
├── mailer.py        Gmail sending, replies, bounces, warm-up
├── schedule.py      prospect time zones and the sending window
├── templates.py     email step templates
├── compliance.py    strict-region flags
├── lint.py          copy checks
└── decks.py         PDF decks
agents/  skills/     the four agents and the Claude Code skills

Credits

Inspired by 4ruple (the local-app, your-own-Claude approach) and by open-source sales and outreach projects. See NOTICE.md.

License

MIT