An AI agent that runs cold email, and sends nothing without approval.
outr
AI outreach platform
- Platforms & AI
- Product
- Brand
From an outreach agency to software anyone can run
outr is an AI agent for B2B cold email. A founder tells it what they sell and who should buy it. The agent finds and verifies the buyers, writes the emails, sends them from warmed inboxes and sorts the replies, and it asks for approval before anything goes out.
It started as Metrichain, a done-for-you outreach agency with a client portal, where a team built and ran every campaign. We rebuilt it from scratch as self-serve software: email only, plans from €159 a month in place of agency packages from €400, and an onboarding where the agent builds the whole campaign before the customer pays.
We designed and engineered all of it: the product strategy, the brand, the landing page and onboarding, the web app, the agent, the sending setup, billing and the admin tools.
Services
- Product strategy
- System architecture
- AI agent
- Web app
- Email infrastructure
- Payments
- Brand
Outreach software asks for trust long before it has earned any.
The people outr sells to have usually tried cold email already: a lead database, a sending tool, a warm-up service, a copywriter, and a domain in the spam folder at the end of it. They are wary of anything that sends email in their name.
So the product had to replace that whole stack with one agent, show a finished campaign before asking for money, and make sure no email reaches a buyer without the founder's click.
Underneath, it had to keep every customer's sending apart, set up and warm inboxes without double orders, run long jobs that survive a timeout, and stop a language model from ever sending a broken email.
The landing page and the onboarding
The live landing page on desktop, from the hero to the pricing, and the onboarding on a phone, where the agent asks one question at a time and answers every pick in its own words.
Plain code where it can be, a model only where it must
The agent works more like an employee than a chatbot. Most of what it does is plain code: routing, frameworks for the emails, rules for sorting replies, maths for the plans. The language model only writes, and every draft passes a check before it can ship.
Around it sits the workspace a founder actually uses: what the agent did today, campaigns with their sequence and funnel, an inbox of replies sorted by intent, the leads, the sending inboxes and the analytics.
What we built
Scope
What we built, and how it works
The full build, area by area: what each part of the product does, and the engineering behind it.
- Landing page and onboarding
- The public site, pricing, a blog, and a four-part onboarding that builds a full campaign before the visitor signs up.
- The onboarding runs in the browser until signup, then carries every answer into the new account, so nothing is asked twice.
- Web app
- Overview, agent workspace, campaigns, inbox, leads, sending inboxes, analytics and settings, on one sidebar.
- Server-rendered pages that read the database directly, with real unread counts and read state in the inbox.
- AI agent
- Request routing, a library of skills with credit costs, two approval gates and compliance guardrails.
- Plain rules route most requests, and a model is used only to write. Every job is saved as a durable task after each step, so long work resumes after a timeout instead of starting again.
- Quality controls
- A check on every model-written email and a reply classifier.
- A draft that fails its checks is retried once, then replaced by a tested template. Replies are sorted by rules before any model sees them.
- Sending
- Separate sending inboxes, set up, warmed and assigned to each customer automatically.
- Inboxes are only used once they reach the inbox reliably. Every order is reserved before it is placed and reconciled after, so a retry never buys twice, and every customer's sending is kept apart.
- Data and security
- Accounts, email and social sign-in, and a database with row-level security on every table.
- Anything that touches billing or ownership is written on the server only. Machine-to-machine endpoints each carry their own secret.
- Billing
- Monthly subscriptions, plan limits and credits for the agent's work.
- One plan definition feeds the pricing page, checkout and every limit. Guards stop double charges and payment events that arrive out of order.
- Operations
- An admin studio, scheduled jobs and monitoring.
- Prompts, templates and the inbox pool change without a deploy. Daily jobs sync stats, enforce limits and write insights, and an alert fires when the share of AI fallbacks rises.
145
commits
~49,000
lines of TypeScript
33
database migrations
3
scheduled jobs
Built with Next.js, React and TypeScript on PostgreSQL.
Try it
What the agent does when you ask
Each quick action runs a fixed plan of steps with credit costs. Launching a campaign stops twice for approval: once before anything is built, and once before anything is sent.
Quick actions
Most of the work is plain code and only the writing goes to a model. Talking to the agent costs nothing; each step that does work spends plan credits.
- Mapping out your campaignfree
Try it
The onboarding, on its own logic
outr's real questions and the agent's real reply to each pick. Choose a buyer and the profile updates, choose a tone and the opener is rewritten, choose a target and the plan maths and benchmark ranges run.
Whose inbox should I land in?
What world are they in?
Where are your buyers?
How big are the companies you sell to?
How should your emails sound?
How many new clients a month would make this a clear win?
Playing sample answers. Pick any option to take over.
Marketing it is. I lead with growth and numbers.
- Role
- Marketing leaders
- Industry
- SaaS & software
- Region
- Germany
- Company size
- 11-50 people
Only people matching all of this get an email. Everyone else is filtered out.
15 minutes?
Lukas - one idea for Northpeak. Worth 15 minutes?
- Lukas MüllerHead of Marketing · Northpeak
- Anna SchmidtCMO · Flowstack
- Felix WagnerMarketing Director · Cadence Labs
75-150
replies / mo
19-35
meetings / mo
4-11
new clients / mo
17 inboxes × 20/day × 22 days = 7,480 emails · 2 per lead. Growth reaches 3,740 verified buyers a month, which puts your target inside the expected range.
Try it
The check every model draft passes
Language models add notes, drop tags and invent variables. Before a draft can reach a buyer, outr strips it and checks it, and if anything fails the tested template goes out instead. Pick a broken draft or edit one.
- Model notes removedThe model's own commentary was cut away.
- Under 1,200 characters178 characters.
- Long dashes replacedA long dash became a plain hyphen.
- First-name tag keptPersonalisation intact.
- Only supported tagsfirst_name, last_name, company_name.
Hi {{first_name}},
We work with founders at SaaS companies in Germany - and {{company_name}} looked like a fit.
If that is on your radar this quarter, worth a quick look?
AlexTry it
The reply classifier
Type any reply and see where it lands. Deferrals like maybe next quarter are checked before interest, so the agent never answers a no with an eager pitch.
Meeting
Wants to talk or meet.
Rules run first and cost nothing on every reply. With AI switched on, a model refines only the unclear ones.
Brand
A calm system for a product that sends email in your name
Navy and white for the work, coral only where a decision is needed. Three typefaces with one job each, and a mascot kept to the onboarding.
outr.
- Grayfel
- landing headings
- Open Runde
- the app
- Geist Mono
- numbers and code
A 3D cat that appears in the onboarding dressed as whichever buyer the customer picks. Everywhere else the product stays plain.
Calm, navy and white for the work, coral only for what needs a decision: the primary button, a hot reply, a number that matters. Every figure in the interface is traceable to a formula or labelled as a benchmark range.
600,000 emails in its first two months
In its first two months, outr sent 600,000 emails, and 3,435 accounts were opened on it.
Every number a buyer sees can be traced: plan capacity comes from the inbox maths, projections are benchmark ranges labelled as ranges, and no email can leave without the customer's approval.
Questions about outr
Next project
Eastframe
Agency website