Why AgentSDR
Outbound usually means stitching together a lead database, an enrichment tool, an email sequencer, a LinkedIn tool, a dialer and a CRM, each with its own seats, its own copy of your leads and its own idea of who replied. Most of them are thin layers over APIs you could call yourself.
- One workspace instead of six tools. One record per person and company, so a reply on LinkedIn shows up in the same queue and pipeline as an email reply.
- The AI does the tracking. Every reply, on every channel, is read, classified and given a drafted answer. You approve what goes out.
- You own it. A single Next.js app and one PostgreSQL database on your own server. Your leads, conversations and recordings never leave infrastructure you control.
- Your keys, no seats. Connect your own Google Workspace, Unipile, storage and any AI model on OpenRouter. No per-seat or per-contact fees: you pay your server and your providers.
Everything in one workspace
Also inside: a lead database with custom fields, CSV and XLSX import with merge fields, Do Not Contact across every channel, organizations with roles, teams and invitations, and per-organization sending rules with safe defaults.
Works with
Get started
You need a machine with Docker.
mkdir agentsdr && cd agentsdr
curl -O https://raw.githubusercontent.com/Kandid-ai/AgentSDR/main/docker-compose.yml
docker compose up -d # PostgreSQL, the app on :3000 (it creates the schema and runs the scheduled jobs)
Open http://localhost:3000: the setup page creates your admin account and organization. Then connect your services in Settings. For a server, replace the # CHANGEME values in docker-compose.yml (or a .env beside it) first; openssl rand -hex 32 makes a secret.
The Docker Compose guide walks through every setting and every service. There are also guides for running from source, deploying on a VPS with Dokploy and going to production.
You need Bun 1.2+ and PostgreSQL 16+.
git clone https://github.com/Kandid-ai/AgentSDR.git && cd AgentSDR
bun install
createdb agentsdr
cp .env.example .env.local # set DATABASE_URL, BETTER_AUTH_SECRET, INTEGRATION_CREDENTIALS_KEY
bun run db:setup # create the schema in the empty database
bun run db:seed:demo # optional: a fictional demo organization
bun run dev # http://localhost:3000
With the demo seed, sign in as [email protected] / demo-password-123.
Environment variables cover only what the app needs before it can read its database. Everything else is per organization, in the app.
| Variable | Required | Purpose |
|---|
DATABASE_URL | yes | PostgreSQL 16+ connection string |
BETTER_AUTH_SECRET | yes | Signs sessions |
BETTER_AUTH_URL | yes | Public URL, read at runtime |
INTEGRATION_CREDENTIALS_KEY | yes | Encrypts integration credentials (AES-256-GCM) |
UNSUBSCRIBE_SECRET | for email | Signs unsubscribe links |
CRON_SECRET, OUTREACH_TICK_SECRET | no | Only to call the job endpoints from an external cron |
RESEND_API_KEY, AUTH_EMAIL_FROM | in production | Auth email through Resend |
AUTH_SIGNUP | no | invite-only (default) or open |
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET | no | "Continue with Google" |
DB_POOL_MAX | no | Connection pool size (default 8) |
Every variable, with how to generate it and where it is read: Configuration reference.
How it works
flowchart LR
A["Leads<br/>import or build a list"] --> B["Tables<br/>enrich with 15 providers and AI"]
B --> C{"Reach out"}
C --> D["Email<br/>your mailboxes"]
C --> E["LinkedIn<br/>invites and messages"]
C --> F["WhatsApp<br/>messages and calls"]
D --> G["Every reply<br/>one queue"]
E --> G
F --> G
G --> H["AI CRM<br/>classify and draft"]
H --> I["You approve<br/>one click"]
H --> J["Pipeline and analytics"]
One Next.js app and one PostgreSQL database: pages, API routes and background workers run in a single process, with no queue server or Redis. Every request, job and webhook runs inside one organization's scope, and code that forgets fails closed. Integration credentials are encrypted per organization and verified with a live call before they are saved. More in How it works and the architecture guide.
flowchart TB
subgraph You["Your infrastructure"]
App["AgentSDR (Next.js)<br/>pages, API, in-process workers and scheduled jobs"]
DB[("PostgreSQL 16+")]
Ext["Call Recorder<br/>Chrome extension"]
end
App <--> DB
Ext -- "recordings" --> App
App <--> Unipile["Unipile<br/>LinkedIn, WhatsApp"]
App <--> Gmail["Google Workspace<br/>Gmail"]
App <--> R2["Cloudflare R2<br/>recordings"]
App <--> OR["OpenRouter<br/>AI models"]
App <--> Enr["Enrichment providers"]
Unipile -- "webhooks" --> App
Gmail -- "Pub/Sub push" --> App
src/app/ Pages (App Router) and API routes
src/lib/<domain>/ Business logic, schema and queries per domain
src/components/ React components, one folder per domain
src/jobs, src/functions LinkedIn sending engine (run by the in-app scheduled jobs)
src/services/ Unipile API clients
db/ schema.sql (generated snapshot) and seed.sql
scripts/ Migration history, audits, database tooling, e2e checks
extensions/whatsapp-recorder/ Chrome extension that dials and records WhatsApp calls
docker/, docker-compose.yml Self-hosting
docs/ Documentation (published at docs.agentsdr.ai)
| Command | What it does |
|---|
bun run dev | Start the app at http://localhost:3000 |
bun run test | 600+ unit tests |
bun run typecheck / bun run lint | Must pass before every commit |
bun run db:setup / db:seed:demo | Create the schema / a fictional demo organization |
bun run db:check | Audit that every table is organization-scoped |
CI runs typecheck, lint, tests, a secret scan, the production build, the extension build and a fresh install on PostgreSQL 16, 17 and 18. See the development guide.
Built with Next.js 16 and React 19, TypeScript, Drizzle ORM, Better Auth, Tailwind CSS with AlignUI, and Bun.
Security and privacy
- Your data stays with you. Leads, conversations and analytics live in your PostgreSQL; call recordings in your own R2 bucket, reached only through short-lived signed links.
- Encrypted credentials. Every saved API key and token is encrypted with AES-256-GCM.
- Tenant isolation, tested. An end-to-end suite signs in as a second organization and proves it can see, change and count nothing of the first.
- Humans approve outbound replies. AI drafts wait for a person, and outbound work can be paused instantly.
- Report vulnerabilities privately through GitHub security advisories. See SECURITY.md.
Responsible use
AgentSDR contacts people on your behalf, so how it is used is your responsibility. LinkedIn's terms do not allow automation and accounts can be restricted; WhatsApp bans numbers that send unsolicited bulk messages; cold email is regulated by laws such as CAN-SPAM, GDPR and CASL; and many places require consent before a call is recorded. Sending limits, unsubscribe links and Do Not Contact reduce the risk but do not make any use compliant. Read Responsible use before you start.
FAQ
Yes. It is open source and you host it yourself, so there are no seats, tiers or per-contact fees. You pay for your server, the accounts you connect (Google Workspace, Unipile) and your own AI usage. It is MIT-licensed, so you can also use, change and build on it commercially.
Google Workspace mailboxes, connected once through a service account with domain-wide delegation. Each mailbox has its own daily limit (30 by default), sending hours and signature. Microsoft 365 is not supported yet.
No tool can promise that, but AgentSDR stays well inside LinkedIn's limits: 30 invitations a day on premium accounts and 5 on free, a random 30 to 60 second gap between invitations, optional working hours per account, and an account that hits LinkedIn's own limit pauses for the day.
You link your number through Unipile and install the AgentSDR Call Recorder Chrome extension. When you call a lead from AgentSDR, the extension dials inside WhatsApp Web and records both sides; the recording goes to your storage bucket and your chosen model transcribes it. It relies on WhatsApp Web's English interface.
Any model on OpenRouter, with your own key. Each call is pinned to the provider you picked, with fallbacks turned off, so your data only goes where you chose.
No. Drafts, including follow-up sequence steps, wait in Action required until someone sends them, as written or edited. The AI can move a lead forward in the pipeline when it is confident; anything else is held for a person.
Yes. Invite teammates from Settings → Members with the owner, admin or member role. One deployment can also hold several organizations, each fully separate.
Community and support
Contributing
Contributions are welcome: bug reports, fixes, enrichment providers, docs. Start with CONTRIBUTING.md for the development setup, the checks a pull request must pass and the commit conventions; issues labelled good first issue are a good way in. Also see the Code of Conduct, how the project is run and what changed.
License
AgentSDR is open source under the MIT License:
- You may use, copy, modify, self-host, distribute and sell AgentSDR, including in commercial products and as a hosted service.
- You don't have to publish your changes, though contributions back are always welcome.
- Keep the copyright and license notice in copies you distribute.
The LICENSE file is the binding text.
