Organic, automated AI marketing that finds its own customers.
Enter your product's website. Eveil reads it, works out who buys it, then markets it for you across every channel a small team can't staff by hand:
- Cold email, the core: finds the companies and the people at them, writes the sequence, sends it from your own mailbox and reads the replies.
- Social posts on LinkedIn, X, Bluesky and Instagram, drafted from what Eveil knows about your product and queued for your approval, with a short video of your product in action when it helps.
- SEO articles for your own blog, on what your buyers search for and your site doesn't answer yet.
- Reddit replies in the threads where your buyers already ask for help.
- Paid ads on your ad accounts, run inside a monthly budget you set: new tests launched, what works kept, what doesn't cut.
All of it reads the same knowledge base, so no channel contradicts another. Lead discovery and outreach are B2B; the content channels work for anyone building an online presence, B2C solo founders included. Paid ads run on the same knowledge base and target profiles, inside a monthly budget you set.
That's automation of the same work a human researcher would do by hand: searching the web for the right companies and people, then writing to them one by one from a real mailbox. It is not a purchased contact list, and it does not send through a shared pool of pre-warmed inboxes on your behalf — every lead is found live, and every mail leaves through the mailbox you connected. Slower to start than buying a list. Honest the whole way through.
Self-hostable, AGPL-3.0, and the free edition has no artificial limits: unlimited mailboxes, unlimited leads, your data on your own machine.
Status: v1 The whole outbound loop works end to end: site analysis, lead discovery, sequences, sending, replies, plus organizations, roles, and pay-as-you-go billing for the cloud edition. LinkedIn posting (personal profile, official API) is in too, X, Bluesky and Instagram posting (Bluesky and Instagram published for you, Instagram with a generated image, X copied and posted by you), and so is Reddit reply drafting - manual-publish only for now, since Reddit is currently blocking new API app registration - and SEO articles for your own blog, plus a public API, an MCP server and CRM webhooks for the rest of your stack. Not built yet: ad management (next up), commenting on someone else's LinkedIn post. See Issues for exactly what's left.
- Reads your site, not a form you fill in. Product, audience, and the reason anybody switches, worked out from the site itself and shown to you before anything gets written. It can also read your github repo for a better understanding of your project.
- Finds the companies and the people, on its own. Segments, search terms, fit scores with the sentence that justifies them, over two independent bundled search engines plus OpenStreetMap, no paid data API required to start. See how discovery finds companies.
- Writes sequences that sound like you. One AI-writing-style box per project (tone, language, banned words) that every generated mail obeys.
- Ask Evie instead of clicking through screens. A chat panel that plans and calls the same agents the app's triggered flows use - "find me 50 dental clinics in Lyon" - and pauses for your approval before anything that spends credits actually runs.
- Sends from your own mailbox. Plain SMTP, no relay, no shared sending domain: what arrives is indistinguishable from something you typed.
- Posts to LinkedIn too, evidence-driven. A knowledge-base fact, a client you just won, or relevant industry news, drafted and queued for your approval before anything goes out - official API, personal profile, no automation of connection requests or messages.
- Posts to X and Bluesky. Short posts from the same evidence, plus your own freshly published articles, one queue for both. Bluesky publishes through its free API with an app password; X you copy and post yourself, since its API is paid per post.
- Posts to Instagram, image included. Every caption comes with an image Eveil generates for it, published to a Business or Creator account through your own free Composio account, so nobody has to create a Meta app.
- Short videos of your product, made for the post. When a post shows a feature in action, Eveil also makes a short video: animated type and your own product screenshots filmed by a camera that zooms in, points and clicks, with a voice over and music if you add an ElevenLabs key. It goes out as a Reel on Instagram and as a native video on LinkedIn and Bluesky, each in its own shape, and is ready to download and attach on X. Pause where something is off, say what to change, and get a new version: no designer, no editing software. Rendered on your own instance with Chromium and ffmpeg.
- Finds Reddit threads worth replying to. Live subreddits and evergreen "best X" threads already ranking on Google, drafted in the thread's own tone. Copy, post it yourself, mark it posted - no OAuth, since Reddit is currently blocking new API app registration.
- Writes SEO articles for your blog. On a feature your site never explains, a competitor comparison, a question buyers ask on Reddit, a client win: always something Eveil actually found. Copy it into your blog, give the URL back, and Eveil won't write it twice.
- Reads and threads replies itself, over IMAP, matched on the mail's own
Message-IDso a reply always attaches to the lead it answers. - A bounce circuit breaker, scoped per mailbox, that pauses sending before a bad batch burns a domain's reputation, not per campaign, per address.
- An API and an MCP server for the rest of your stack. Organization tokens to read companies, contacts and campaigns into a CRM, push contacts in, enrol them, or ask for a social post. An AI assistant can do the same over MCP, connecting with a token or by signing in, across all your projects. See the API docs.
- Webhooks for your CRM. A signed request the moment a reply comes in or a deal changes status, per project, with ready-made n8n recipes for HubSpot, Pipedrive and the rest, both ways. See Webhooks and CRM sync.
- Self-hosted and AGPL-3.0. Four containers, five minutes, your data never leaves your machine unless you choose the cloud edition.
- No dark patterns. No open-tracking pixel, no mailbox warm-up, no OAuth lock-in, no purchased contact database. See what it deliberately does not do.
Made by Eveil for one of its own Instagram posts, then refined three times in the editor just by saying what to change ("add a voice and music", "use the real logo"). A director agent invents each video from the post, your brand and your product screens: your colours, fonts and logo read from your site, its own screenshots of the site when you have added none, illustrations it draws and animates itself, and a camera that films your screens, zooming in, pointing and clicking. With an ElevenLabs key it adds a voice over and music. Each network gets its own shape (a 9:16 Reel on Instagram and Bluesky, 4:5 on LinkedIn, 16:9 for X), and you pick whether a draft goes out with its image or its video. It renders on your own instance with Chromium and ffmpeg: no video API, no per-minute fee.
Two ways to run it, same code, no feature gate between them.
Self-hosted. Free, forever, AGPL-3.0. No per-seat or per-mailbox fee. Bring your own AI provider key (Anthropic, OpenAI, whichever you already pay for) and your data never leaves your machine. Organizations, roles and multi-user access are core to both editions, not held back for cloud. Your install starts with limited knowledge of the web and grows smarter on its own from there.
Cloud. Managed hosting at eveil.cloud, with pay-as-you-go credits instead of an AI provider account: top up whatever amount you choose, spend it at one flat published rate, no plans and no subscription. It's smarter from day one, too: the search-host registry and the page-cache discovery it draws on grow with every customer, so a new account skips the cold start a fresh self-hosted install has to work through alone. Cloud does not unlock features; it removes setup, hosting and the AI-key requirement.
1,000 credits is about $1 of AI cost, and SMTP sending, IMAP reading and email verification cost nothing: most competitors bill verification. A full 100-lead campaign runs about 3,500 credits end to end. New accounts start with a trial grant, capped to one project and to leads discovered rather than just credits spent, with no CSV export before a first payment. The whole grid, what every AI step costs and what costs nothing, is published at eveil.cloud/pricing.
You need Docker and about five minutes. No clone, nothing to compile: the app runs a prebuilt image, and three files in an empty directory are the whole install.
mkdir -p eveil/docker/searxng && cd eveil
base=https://raw.githubusercontent.com/Dricle/eveil/main
curl -fsSL -o compose.deploy.yaml $base/compose.deploy.yaml
curl -fsSL -o .env $base/deploy/.env.example
curl -fsSL -o docker/searxng/settings.yml $base/docker/searxng/settings.ymlThree values to fill in, three more you can leave alone:
| Variable | What it is |
|---|---|
APP_URL |
The address people actually type. Every link the app generates is built from it. |
DB_PASSWORD |
Any long random string. The stack creates the database with it. |
SEARXNG_SECRET |
Any long random string. Only signs the bundled search engine's own requests. |
ADMIN_EMAIL, ADMIN_PASSWORD |
Optional. Set them and the first account is created on boot; leave them empty and the setup screen asks instead. |
APP_PORT |
Optional, defaults to 80. The host port the app is published on. Change it if something already holds 80. |
Leave APP_KEY and CREDENTIALS_KEY empty: they are generated on first boot.
Fill them in only if you would rather manage them yourself; anything set there
wins.
Leave the MAIL_* block for later too: the app boots without it, but
password resets, member invitations and email verification need it filled in
before they can actually send. See Configuration.
Then bring it up:
docker compose -f compose.deploy.yaml pull
docker compose -f compose.deploy.yaml up -dOpen APP_URL. You land on the setup screen, or on the login if you set the
ADMIN_* pair. Migrations run on boot, before anything serves a request or picks
up a job.
The app speaks plain HTTP: TLS, a reverse proxy or a tunnel, what actually runs
in the four containers, the app's own transactional mail, and where the
encryption keys live — the full Installation
and Configuration docs cover
all of that. -f compose.deploy.yaml on every command gets old fast; put
COMPOSE_FILE=compose.deploy.yaml in .env and drop it.
Prefer to compile what you run? Clone the repository and add
compose.build.yaml as a second -f:
docker compose -f compose.deploy.yaml -f compose.build.yaml up -d --build.
docker compose -f compose.deploy.yaml pull
docker compose -f compose.deploy.yaml up -dMigrations run on boot; your database, Redis and storage volumes are untouched. See Updating for rolling back, and Backup before you do.
Sign in and give it the address of your product. That is the whole setup: the rest is a guided run you watch and agree with:
- It reads your site. A minute or two, page count ticking up on screen.
- It shows you what it understood: what the product does, who it is for, why anybody switches. Agree, or correct it first. Every mail is written from this, so correcting it now is cheaper than correcting it in three hundred mails.
- Agreeing starts the segments: who buys it, and the search terms that find each one.
- Agreeing to those starts the searching, one search per segment you left switched on. Companies appear in Leads with a fit score and the sentence that justifies it.
Nothing is written to anybody during any of that. Sequences arrive as drafts and nothing sends until you activate a campaign.
Under Settings, Project, there is a box for how the AI writes: tone, language, words never to use. It starts with one rule already in it, banning dash punctuation, because that is one of the cheapest tells that a machine wrote a sentence. Everything whose output you read as prose obeys it, from the product portrait to the opening line of a mail. Edit it or empty it as you like.
Two things the app will nag you about at the top of every screen, because neither announces itself otherwise:
- No AI provider key (shown to the instance's superadmin only, since nobody else can fix it). Without one every agent fails in the queue. Settings → App settings → Provider. The "test" button says exactly why a key was refused, which beats finding out from a job that dies an hour later.
- No mailbox on this project. Everything up to writing a sequence works, but a campaign will activate and then sit there. Settings → Mailboxes: plain SMTP and IMAP, no OAuth, with presets for Infomaniak, OVH, Gandi, Zoho, Gmail and Microsoft 365 — see Configuration for the Gmail/Workspace and Microsoft 365 connection quirks, and for testing the whole loop (search → sequence → send → reply) against your own mailbox before you point it at real leads.
The development stack is Laravel Sail, at the root of the repository:
cp .env.example .env
./vendor/bin/sail up -d
./vendor/bin/sail artisan key:generate
./vendor/bin/sail artisan eveil:credentials-key
./vendor/bin/sail artisan migrate
yarn install && yarn devHost ports are shifted so this runs beside other projects: app on 8080, Postgres on 5442, Redis on 6382, Mailpit on 8035.
PHP runs in the container, JS tooling runs on the host. node_modules is
installed with macOS binaries and mounted into a Linux container, so eslint and
Vite fail inside Sail; and the host PHP must be 8.4+, because yarn build shells
out to php artisan wayfinder:generate. For the same reason, never run
wayfinder:generate inside the container. It emits route modules without the
.form() helper and the type check then fails in files you did not touch.
./vendor/bin/sail artisan test # the suite
./vendor/bin/sail composer lint # Pint
yarn lint:check && yarn types:check # eslint, vue-tscPostgres is required, including for tests: the schema leans on JSONB, partial
unique indexes and ilike, so SQLite would pass tests that production fails.
Two documents are worth reading before changing anything.
GUIDELINES.md holds the decisions and the reasoning behind them; .ai/rules/
holds the settled conventions and the traps somebody already walked into.
- No open tracking. No pixel, no link rewriting. Apple's Mail Privacy Protection and Gmail's image proxy make open counts fiction, and the pixel costs inbox placement. The metric here is the reply.
- No mailbox warm-up. Warm-up serves high volume from fresh domains, which is not what this is for, and shared warm-up networks are increasingly a negative signal to Google and Microsoft.
- No OAuth. SMTP and IMAP credentials only.
- No unsubscribe link. Nobody subscribed to anything, so an unsubscribe button contradicts a hand-written message. The opt-out is a sentence in the body, and a reply asking to stop suppresses the address immediately.
- No purchased contact database. Every lead here was found and read.
AGPL-3.0. If you run a modified version as a service, you have to publish your changes.