Apps and AI agents that keep working after we leave.

Forever Works is a senior engineering firm. We build web and mobile products in weeks and put AI agents to work in the systems you already run: your inbox, your invoicing, your documents, your customers. You own all of it from the first day.

Book a 30-minute call See what we would build for you

What do you need?

Pick the closest. Each one opens the first draft of how we would do it, what it takes, and how you will know it works.

What do you need?

An agent that does one job in the tools your team already uses.

Triage the inbox, draft and issue invoices, match bank statements to receipts, file documents, answer customers. Each agent works through a narrow connection to the system it acts in, every action is logged, and anything that moves money or files with an authority waits for a person to approve it.

  1. Week 1We sit with the person who does the job today, write down how they decide, and connect read-only to the systems involved.
  2. Weeks 2 to 4The agent works next to them. It drafts, a person approves, and every decision goes into the test set.
  3. AfterWhat the agent gets right every time stops needing approval. Money and filings keep their approval gate unless you decide otherwise.

How you will know it works

An eval set made from your own cases, run on every change, with the score in front of you. Our invoicing agent ran 24 tasks, including prompt-injection attempts, with 0 writes nobody approved.

Built like it: Our own back officeAgent evals

A web app, an iOS and Android app, or all three from one codebase.

React, React Native and Expo, TypeScript end to end, with the backend, the sign-in, the payments and the store listings. Designed screen by screen with you, and released in small steps you can try on your own phone from the first week.

  1. Week 1A working build on your phone, made from your real screens and your real data.
  2. Weeks 2 to 6The product, released every few days. You use it, we change it.
  3. LaunchApp Store and Google Play review handled, the web version on your domain, crash reports on.

How you will know it works

Every flow runs on iOS simulators and Android emulators on every change, and you get the screenshots. A release candidate goes to your own phone through TestFlight and Google Play testing.

Built like it: RitualTilia Studio

Search and answers over your documents, with the source under every sentence.

Policies, contracts, manuals, tickets, wikis. The service answers from your documents, cites the passage it used for each sentence, and says so plainly when the answer is not in them. It runs on the model you choose, in the cloud or on your own servers.

  1. Week 1We index a sample and write 40 real questions with you, each with the right answer and where it lives.
  2. Weeks 2 to 3Retrieval and prompts tuned until the questions pass the bar we agreed before starting.
  3. AfterIt goes where your people already ask: Slack, your intranet, your product, or a chat page of its own.

How you will know it works

Your questions become the test, and you see the score before and after every change. On our own 40-question bilingual set, hybrid retrieval found the right passage in the top five 82.5% of the time, up from 52.5% with keyword search.

Built like it: Ask the handbookAgent evals

An MCP server for your product, so assistants can use it on your customers' behalf.

Your customers increasingly ask Claude, ChatGPT and their own agents to do the work in your product. An MCP server gives those assistants a safe, documented way in: sign-in and consent, scopes, rate limits, and tool descriptions a model actually follows.

  1. Week 1The three things your customers ask for most, working from Claude and ChatGPT against your staging API.
  2. Weeks 2 to 3OAuth, scopes and limits, tests against the models your customers use, and the cases where a model would misuse a tool.
  3. AfterPublished, documented and listed, with usage you can see.

How you will know it works

We build and run these every day: our own company works through 21 MCP servers we wrote, and fifteen of them are published on npm and in the official MCP Registry.

Built like it: MCP serversOur own back office

Take an AI feature from demo to production.

We take an AI feature that impressed in a demo and make it dependable: evaluations on your real traffic, guardrails, prompt-injection testing, cost and latency budgets, monitoring, and a security review of what the model can see and do.

  1. Week 1We measure it as it is, on your cases, and agree the pass mark with you before changing anything.
  2. Weeks 2 to 4Fixes, each one measured against the same set, so you see what helped and what did not.
  3. ShipIt goes live when it passes, with the dashboard and alerts that keep it honest after launch.

How you will know it works

A results table you can read, before and after, on cases from your own users. On our agent evals, retrieving before acting raised the pass rate from 21 of 24 tasks to 24 of 24.

Built like it: Agent evalsOrb

A senior engineer who joins your team, works in your stack, and ships.

For teams that have the plan and need the hands. The engineer you meet on the first call is the one who does the work, in your repositories, your tickets and your stand-ups, for a fixed number of days a week.

  1. Week 1Access, the codebase walked through with your team, and a first change merged.
  2. Every weekMerged work you can review, and a short written note of what changed and why.
  3. The endA handover your team can run without us, on a date agreed at the start.

How you will know it works

Three years as enterprise architect of a governance platform at a tier-1 bank, from whiteboard to production; a platform rebuild for a media-tech company, delivered and handed over inside a fixed engagement.

Built like it: OrbMeridian

You own it from the first day.

We plan the handover on the first day and write its date into the proposal. Your team can run what we built without us, and keeping us on to run it is a choice you make each month.

  • The code, in your repositoryUnder your account from the first commit, with its history.
  • The prompts, the evals and the test casesNext to the code, so any engineer can see what changed and whether it helped.
  • Every credential in your own vaultOur access is granted by you, scoped, and removed on the handover date.
  • Runbooks your team can followHow to deploy it, watch it, roll it back and change the model behind it.
  • A recorded walkthroughAn hour on how it fits together, for the people who will own it.
  • Your choice of modelAnd the tests that let you switch to another one safely.

Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendors' forward-deployed engineering teams, “trapped by soaring costs and unable to evolve it on their own.” Gartner, 29 September 2026. Its advice: settle IP ownership, knowledge transfer and an exit strategy from day one.

Work you can check.

Dates from the commit history, tests and scores from the test runs, and live links where the thing is public.

  1. Ritual

    Medication app for iPhone and Android, in beta

    One tap records the usual dose, reminders arrive at your hour, and sync is end-to-end encrypted. Native widgets, Apple Watch and Wear OS, and a catalogue of 1.59 million medicines from 92 public sources.

    ritual.day →

    First commit
    4 Aug 2026
    On Android test track
    7 Aug 2026, day 3
    Tested by
    49 Maestro flows on iOS simulators and Android emulators, about 1,567 test cases
    Status
    Beta: TestFlight and Google Play testing
    Ritual on an iPhone simulator: History, 35 records across 7 days in a chart by time of day

  2. Ask the handbook

    Document chat with citations

    Ask a company handbook a question and every sentence of the answer cites the passage it came from. Questions the handbook cannot answer are turned away before the model runs. Llama 3.3 70B on Cloudflare Workers AI, with fictional data so anyone can try it.

    Try the handbook →How it is tested →

    Built
    17 Oct 2026, in one night
    Eval
    25 of 25 questions (first run 23 of 25)
    Model
    Llama 3.3 70B, bge embeddings
    Ask the handbook answering how much a hotel in Paris may cost, with the cited passage highlighted beside the answer

  3. MCP servers

    Open source, MIT

    The connectors our own agents use, published for anyone: Google Workspace, Notion, PSD2 banking, Romanian e-invoicing and tax, invoicing, domains and DNS, Expo builds. Fifteen are on npm and listed in the official MCP Registry.

    github.com/foreverworks/mcps →

    Servers
    21, with 451 tools
    Tests
    about 2,584 cases in 224 files
    Published
    npm and the MCP Registry, Oct 2026

  4. Agent evals

    Test harness for an invoicing agent

    An agent that issues invoices, tested on 24 tasks across reads, writes and prompt-injection attempts, with its approval gate kept outside the model, where a prompt cannot reach it. Its retrieval was measured on a 40-question bilingual set.

    Right passage in top 5
    52.5% keyword, 82.5% hybrid
    Tasks passed
    21 of 24 acting first, 24 of 24 retrieving first
    Unapproved writes
    0

  5. Orb

    Governance platform at a tier-1 bank

    A declarative engine that governs cloud environments, GitHub organisations and access control across AWS, GCP and GitHub, designed from the whiteboard with the bank's security, platform and engineering teams. Its security layer, from authentication and key pairs to an internal certificate authority, runs in production.

    Client
    A tier-1 bank
    Role
    Enterprise architect
    Years
    2023 to 2026
    Runs on
    Terraform across AWS, GCP and GitHub Enterprise

All work, with the full records →

Our own company runs on the agents we build.

The agents work on our own servers through connectors we wrote: one per system, each running as its own locked-down service with its own credentials. They act on a person's request or on a schedule a person set. Sending mail in our name, public replies and purchases wait for that person's approval, and every session leaves a full transcript.

21connectors we wrote and run, with 451 tools between them
35agent sessions at once, measured on 28 September
1,561commits to the fleet's own setup since July; deploys reach a test machine before production
6machines, from a build server to an iOS simulator host
  • BankReads the business account over PSD2 and matches each transaction to its receipt or invoice. It cannot move money.
  • e-InvoicingValidates and files e-invoices and transport declarations with ANAF, and follows each one to its answer.
  • InvoicingIssues invoices and records the payments against them.
  • Tax accountKeeps every document the tax office posts and checks declarations with ANAF's own validator.
  • Mail and calendarTriages, drafts and schedules across two Google accounts.Sending waits for a person
  • Mobile releasesBuilds the apps, runs the test flows on iOS simulators and Android emulators, reads tester feedback.Store replies wait for a person
  • DomainsDNS records and renewals.Purchases wait for a person
  • Our own documentsSearch with local embeddings and a reranker, on our own server.

Who you work with.

Horațiu leads every engagement. Larger projects add senior engineers he has vetted on real work; each one is named in the proposal, and you meet them before you sign.

Horațiu Vlad

Founder and lead engineer

Thirteen years building platforms where a mistake is expensive: banking, payments, industrial software. He is the person on your first call, and the person who answers for the work after it.

He has worked for Deutsche Bank, Global Payments, Siemens Industry Software and Mediatel Data.

  1. Forever Works. AI agents, retrieval and MCP integrations in production, each with its tests and evals.
  2. A media-tech platform. Sole architect of a platform rebuild on AWS, with the operations tooling and the handover.
  3. A tier-1 bank. Enterprise architect.
  4. Mediatel Data. Software architect of a cloud move for a call-centre platform with 20M+ monthly users.
  5. Global Payments. Senior full-stack engineer on Heartland point-of-sale systems.
  6. Siemens Industry Software. Full-stack engineer on Simcenter Amesim Cloud.

Questions buyers ask us.

What does it cost?

Every phase has a fixed price. We quote it after the first call, in writing, before you commit to anything. A first phase is usually one to four weeks.

Which AI models do you use?

Whichever fits the job and your rules on data: Claude, GPT, Gemini, Llama, Mistral, or an open model on your own servers. We resell none of them and earn nothing from your choice, and the evals we leave behind let you switch later.

Where does our data go?

It stays in your accounts and goes only to the model provider you approve, or nowhere outside when the model runs on your own servers. We sign your NDA and data processing agreement before we see anything.

Where are you, and how do we work together?

Brașov, Romania, inside the EU, on Central European Time. We work in English and Romanian, remotely or at your office, with a written update every week.

Who actually does the work?

Senior engineers only. Everyone who works on your project is named in the proposal, and you meet them before you sign.

Tell us what has to work.

Book a 30-minute call hello@foreverworks.com