HeyLead — AI agent for LinkedIn outreach

HeyLead

HeyLead is an AI agent for LinkedIn outreach: it finds the right people, writes to them in the voice of your own LinkedIn posts, follows up, and handles replies. It runs from Claude Code, Cursor, any MCP client or a web dashboard. Sign in on the web, then talk to your AI and say "find me leads."

Use cases

A campaign has one of six goals, and the ICP, the fit check and the messages follow it: Sell a product or service, Find a job, Hire people, Find partners or investors, Find a vendor and Research interviews.

• Sell a product or service: Reach the people who buy what you built.
• Find a job: Reach the people who hire for the role you want.
• Hire people: Reach candidates for a role you are filling. Works with a custom brief.
• Find partners or investors: Reach the people who can sign a partnership or an investment. Works with a custom brief.
• Find a vendor: You are the buyer. Reach the people who sell what you need.
• Research interviews: Reach people to interview, survey or test with. Works with a custom brief.

Hire people, Find partners or investors and Research interviews run on a custom brief until their message sets exist.

Getting Started

MCP (Model Context Protocol) lets AI assistants use external tools. HeyLead gives your AI the ability to do LinkedIn outreach for you.

You need: Cursor or Claude Code — any MCP-compatible AI editor.

Step 1: Install HeyLead

Claude Code, as a plugin (hosted, nothing to install):

/plugin marketplace add D4umak/linkedin-outreach-mcp/plugin install heylead@heylead

Then run /mcp, pick heylead and choose Authenticate to sign in to your HeyLead account. The plugin connects the hosted server at https://heylead.dev/mcp and adds skills for starting a campaign, approving waiting messages, answering replies, a weekly review, pausing, and adding a list of people. Its source is in plugins/heylead.

Or run the client locally over stdio. Install the CLI once with uv:

uv tool install heylead

Claude Code:

claude mcp add heylead -- heylead

Cursor: Settings > MCP > "Add new MCP server" > Name: heylead, Command: heylead

Any MCP client:

{  "heylead": {    "command": "heylead",    "args": []  }}

Update with uv tool upgrade heylead.

Step 2: Set up your account

Option A — Hosted (easiest): take the 90-second quiz or sign in at heylead.dev/dashboard/login. Your quiz personas become a draft campaign. Connect LinkedIn in Settings → Connected accounts, then launch from the dashboard — or copy your setup message from Settings → Integrations → Chat client into your AI chat. Nothing sends before you launch. Hosted users share a professional directory; campaigns and inboxes stay private.

Option B — Self-hosted: run everything against your own accounts. You need two things first:

1. A Unipile account — this is what talks to LinkedIn. Sign up at unipile.com, then put the DSN and API key from the Access Tokens page, then run heylead config set unipile_api_url <DSN> and heylead secrets set unipile_api_key (it prompts; the key is kept in your OS keychain, never in the config file).
2. An LLM API key — AI calls are billed to you. A free Gemini key is enough to start.

Then open your AI chat and say:

"Set up my HeyLead profile with this Gemini key: YOUR_KEY"

You'll get a LinkedIn authentication link. Open it, connect LinkedIn, then say "finish setup". HeyLead fetches your profile and analyses your writing style.

Step 3: Find leads

"Find me CTOs at fintech startups in New York""Send outreach to the campaign""Check my replies""How's my outreach doing?"

How It Works

1. Define your ICP — "Generate an ICP for AI SaaS founders" → RAG-powered personas with pain points, barriers, and LinkedIn targeting
2. Create a campaign — "Find me fintech CTOs" → searches LinkedIn, scores prospects by fit
3. Warm up prospects — Engages with their posts (comments, likes) before reaching out
4. Send personalized invitations — Voice-matched messages that sound like you, not a bot
5. Follow up automatically — Multi-touch sequences after connections are accepted
6. Handle replies — Detects sentiment, advances positive leads toward meetings, answers questions
7. Track outcomes — Won/lost/opted-out tracking with conversion analytics

Sending model: campaigns are created as drafts and only start when you explicitly launch them. Every send passes rate limits, working-hours checks, and a 1st-degree connection guard before it goes out. HeyLead sends from your own LinkedIn account at a human pace: at most 20 invitations a day and 100 a week on a free LinkedIn account (more on Premium or Sales Navigator: a newly connected paid account starts at 20 a day and doubles each week up to its limit), Monday to Friday 08:00 to 22:00 in your time zone, minutes apart. It backs off when LinkedIn pushes back and resumes on its own. You can pause any campaign at any time. Launching is also what commissions cloud sending — in observe mode nothing is commissioned and nothing is sent, from either machine.

Approval mode (hosted): a workspace that has not chosen otherwise holds every opening DM and follow-up until a person approves it, and the approved text is what is sent. Read what is waiting on the dashboard's Approvals page or with inspect(action="waiting"); decide with prospect(action="approve_message") or discard_message. Autopilot sends them unread and is one switch away (Settings → Sending, or scheduler(action="approval_mode", mode="autopilot")). Invitations, InMail and replies are never held.

Tools

HeyLead gives your AI 22 tools:

Core Workflow

Outreach & Engagement

Campaign Management

Insights & Analytics

Growth & Relationships

Automation & Account

Key Features

Voice Matching — Analyzes your LinkedIn profile and posts to capture your writing style. Every message sounds like you wrote it.

ICP Generation — RAG-powered pipeline that crawls company context, generates buyer personas with pain points, fears, barriers, and maps them to LinkedIn search parameters.

Autonomous Scheduler — Runs in the background, respects working hours and rate limits. On a hosted account, cloud is the default sender for every campaign. Launching commissions the cloud, so outreach continues with your laptop closed, inside the sending window: invitations, opening DMs, first-touch InMail, follow-ups, engagements, follows, endorsements, email fallbacks, prospect top-ups, auto-replies, inbound, warmup, signal collectors, and post-intel. This machine does not start a local scheduler engine for that work. Move the whole account here with scheduler(action='send_from', host='local'), which turns the cloud scheduler off. Observe still means nobody sends. Direct / self-hosted installs send from this machine only.

Engagement Warm-ups — Automatically engages with prospect posts before sending connection requests, building familiarity.

Pace — HeyLead sends from your own LinkedIn account at a human pace: at most 20 invitations a day and 100 a week on a free LinkedIn account (more on Premium or Sales Navigator: a newly connected paid account starts at 20 a day and doubles each week up to its limit), Monday to Friday 08:00 to 22:00 in your time zone, minutes apart. It backs off when LinkedIn pushes back and resumes on its own. You can pause any campaign at any time.

Outcome Tracking — Mark deals as won/lost, track conversion rates, identify stale leads, measure engagement ROI.

Pricing

A connected account is a LinkedIn seat or an email mailbox. Free includes one of either.

Invitation limits follow the LinkedIn account (free, Premium or Sales Navigator), not the HeyLead plan. Self-hosted free installs have monthly quotas: 50 invitations, 20 messages, 30 engagements, 1 active campaign, 3 ICP generations.

Privacy

• AI calls — routed through HeyLead's backend or your own key
• Cloud MCP — your data is processed server-side but never shared with third parties
• Local mode — contacts and messages stay on your machine in a local SQLite database
• Tool-call telemetry — when you are signed in, your HeyLead workspace records each tool call: the tool, whether it worked, the error type, how long it took and the app that called it (Claude, ChatGPT, Cursor…). Never the arguments, the results, a prospect's name, a link or a message. Turn it off with heylead config telemetry off or DO_NOT_TRACK=1

Power users: Pass your own LLM key (Gemini/Claude/OpenAI) during setup to use your own AI. Completely optional.

Where secrets are kept

HeyLead keeps its backend token, the Unipile key and any LLM keys in the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service), or in an owner-only secrets.json in the HeyLead folder when none exists. The config file holds ordinary settings only. To inspect an install without seeing a secret:

• heylead config get backend_url prints one ordinary setting
• heylead api get /api/v1/campaigns makes an authenticated read of the hosted API, credentials masked
• heylead secrets status lists which secrets are stored, never their values
• heylead secrets clean moves secrets out of an older config file, drops keys a cloud-sending install does not use, and deletes old config backup copies

Backend mode & env

When the MCP client talks to a HeyLead backend (e.g. heylead-api), the backend uses these environment variables. Operators running their own backend should set them as required.

For full backend configuration and deployment, see the heylead-api (or backend) repo and its docs.

Optional Dependencies

The base install covers all core features. For advanced ICP generation:

pip install heylead[icp]    # Embeddings for RAG-powered ICP generationpip install heylead[crawl]  # Web crawling for company context ingestionpip install heylead[all]    # Both

Troubleshooting

"heylead: command not found" or "uv: command not found" Install uv first (curl -LsSf https://astral.sh/uv/install.sh | sh or brew install uv on Mac), then run uv tool install heylead and uv tool update-shell (or pass the full path from command -v heylead).

"MCP server not connecting" Restart your editor after adding the MCP server. In Cursor, check Settings > MCP — the server should show a green dot.

"Setup failed" or "LinkedIn not connected" Make sure you clicked "Connect" on the LinkedIn row of the sign-in page (dashboard: Settings → Connected accounts) and completed the LinkedIn login. Then run setup again.

Need help? Open an issue.

Publishing to PyPI (maintainers)

To make HeyLead available on PyPI (or to publish a new version):

Option A: Publish via GitHub Release (recommended)

1. One-time: Create a PyPI account and an API token. In your repo: Settings → Secrets and variables → Actions → add secret PYPI_TOKEN with the token value.
2. Bump version in pyproject.toml (version = "0.2.4").
3. Commit, push, then create a GitHub Release (tag e.g. v0.2.4, release title optional). The workflow .github/workflows/publish.yml runs on release and publishes to PyPI.

Option B: Publish manually

pip install build twinepython -m build          # creates dist/twine check dist/*       # optional: validatetwine upload dist/*      # prompts for PyPI username + password (use __token__ and your API token)

After publishing, anyone can install it with uv tool install heylead and add it with claude mcp add heylead -- heylead (Cursor: command heylead).

For AI Agents

HeyLead is designed as an MCP-native tool — built for AI agents, not humans clicking buttons.

Install as MCP server (stdio):

{  "heylead": {    "command": "heylead",    "args": []  }}

OpenClaw: Add the same entry to your openclaw.json under mcp.servers. Also available on ClawHub — search "HeyLead".

Sign in at heylead.dev (hosted), or bring your own Unipile account and LLM API key (self-hosted).

Capabilities: LinkedIn lead generation, cold outreach automation, ICP generation with buyer personas, voice-matched personalized messaging, multi-touch drip sequences, reply sentiment classification, engagement warm-ups, campaign analytics, and scheduled sending from the cloud with your laptop closed.

22 tools, from finding the right people to answering their replies: prospect discovery, ICP generation with buyer personas, connection invitations with a note, opening messages and follow-ups, warm-ups (profile views, follows, comments), reply handling and outcome tracking. Replies are read, classified and answered from the campaign's facts. A reply that could mean two things gets one short question back. A question your facts can't answer is held for you.

See AGENTS.md for the full agent integration guide.

Links

• PyPI
• MCP Registry
• Smithery
• ClawHub (OpenClaw skill store)
• Issues

License

MIT (code) — see LICENSE

Knowledge base and prompt configurations are proprietary.