#160inboxyes/noopenansweredprocessingprocessed12m
Tomorrow’s 9:00 clashes with the school run. Move it to 11:30 and tell all three attendees?
Everyone’s calendars are free at 11:30. Drafted notes attached; nothing sends until you say so.
Your own governed AI fleet.
A team of AI agents that runs your inbox, outreach, and operations, and nothing they do ships without your yes.
Talk to us
Or read how it works in the docs.
#161financeyes/no1h
Invoice #2041 is 38 days overdue. Send the reminder tonight and add the late fee from the contract?
#160 · inbox · meeting moved, three notified · approved by you · 14:02:31
#164financeQ2 books sentapproved by you09:41:12
The approval queue: proposals wait here until you decide.
The operational tax
Running your own operation comes with a tax: the inbox, the follow-ups, the invoices, the posting. AI agents can do most of it already, and you haven’t handed it over because nobody wants an unsupervised agent near their email, their clients, or their money. So agents today are either fenced off and useless, or loose and wreaking havoc.
Havenware is the other option: a government for your agents.
This can be your fleet
A fleet, not a chatbot. Build yours within Havenware.
Heron
inboxNotices a thread going quiet two days before you would.
Triages what arrives, drafts your replies, and chases follow-ups before they go stale.
Reply to the invoice query? · est ~1 minAtlas
outreachTreats silence as a scheduling problem.
Researches leads, keeps sequences moving, and updates the CRM so you never have to.
Send the follow-up to the quiet prospect? · est ~2 minVela
socialDrafts in your voice, always.
Drafts posts in your voice and keeps the schedule you always mean to keep.
Post the draft thread? · est ~1 minEcho
financeNever lets a transaction go uncategorized.
Prepares invoices, matches payments, and readies the books for your accountant.
Categorize 2 transactions? · est ~2 minNova
researchWatches the competition and keeps you a step ahead of it.
Monitors the sources you care about, sends digests, and answers questions from your own documents.
Send the weekly digest? · est ~1 min
Your agent's DNA
We wrap your agents in a governance layer that keeps them efficient and safe.
01 · dna
limits: none
A domain, a goal, access to credentials and tools.
02 · the fitting
+ credentials · scoped per agent
+ approval gate · in the execution path
+ audit trail · every action lands
+ secret scanning · keys stay out of transcripts
- unsupervised sends
- unscoped access
The harness control center lets you set the rules on how your agents behave.
03 · governed
limits: enforced
The same agent, governed. The limits travel with it.
Anywhere you are
Your agents meet you where you are: WhatsApp, Discord, Telegram, SMS, email, or a call. Answer from any of them; the decision still lands in one queue and one trail.
your fleet
whatsapptelegramemaildiscordsmscallyour fleet
whatsapptelegramemaildiscordsmscall
atlasagentdirect messagewhatever happened with the Vienna people?
The CFO didn't respond after we sent the deck on May 12. Their new Head of Operations started Monday, ex-Siemens. I've written a draft in your voice to bump that thread, scheduled for tomorrow at 11:00. I'll also ping him on WhatsApp. Approve?
17:28Scheduled for 11:00. The WhatsApp ping is off. It's in the trail.
Governance
You stay in control
One loop keeps the fleet in order: propose, review, approve, execute, audit.
enforced
Every action is a proposal until you approve it. Enforced in code, not policy.
recorded
A full audit trail: which agent proposed what, when, and who approved it.
- 01
Propose
An agent packages the action as a proposal: one question, the context to decide it, an honest estimate of your time.
new proposalfinancequestionOvernight: 14 invoices matched to payments and filed. One supplier invoice is €480 over the quote, with no purchase order attached. Query the supplier about the difference?whyThe 14 matches are in the books already. The query is drafted, and it goes nowhere until you approve it.expiryin 48hest~1 min01 / 05 - 02
Review
The proposal waits in your queue. Nothing happens while it waits.
approval queue1 open#167financeyes/noopen6hexpires in 48h · est ~1 min
02 / 05 - 03
Approve
You answer in a tap, or send it back.
#167financeyes/no6hOvernight: 14 invoices matched to payments and filed. One supplier invoice is €480 over the quote, with no purchase order attached. Query the supplier about the difference?
03 / 05 - 04
Execute
The agent acts, within exactly the scope you approved.
execution log#167financesending query€480 over quote, no PO04 / 05 - 05
Audit
The proposal, your decision, and the action land in the trail.
audit trail#167finance14 matched and filedquery sentapproved by you08:12:44#165inboxdeclined 8:00, offered 11:00approved by you07:32:1805 / 05
Oversight, not babysitting
Human oversight decays when reviewing costs too much: the queue grows and review turns into rubber-stamping. So every proposal is one decidable question, with context attached and an honest estimate in minutes, and agents earn autonomy through clean runs. An agent that proposes well gets a longer leash; one that slips goes back to review.
How it’s built
The primitives, for the technical reader.
- governance
Governance layer
Human sign-off enforced at the code level. Consequential actions require an approval record to execute, not a policy asking agents to behave.
- memory
Per-agent memory
Each agent keeps its own memory of its domain, its history, and your corrections.
- tools
MCP tool access
Agents reach email, calendar, CRM, and finance tools through MCP servers, with per-agent credentials.
- context
Shared knowledge graph
Context is shared across the fleet through a knowledge graph, so what one agent learns, the others can use.
- substrate
Provider-interoperable substrate
Runs on Claude today, built against a provider-neutral interface. Havenware is the harness, not the model: bring your own providers and keys.
- attribution
Full action attribution
Every action traces to an agent, an approval, and a timestamp.
Works with your stack
Your agents reach the tools you already use through MCP, the open standard, so the fleet plugs into your operation instead of replacing it.
- Slack
- Discord
- Telegram
- Gmail
- Google Drive
- Google Calendar
- Notion
- Salesforce
- HubSpot
- GitHub
- Linear
and many more
Your knowledge, compounding
Every correction you make and every approval you grant lands in a knowledge base that is yours: how you write, what you charge, who matters, what good looks like. A containment protocol keeps each part scoped, so the fleet draws on what it needs and nothing else.
Each cluster stays scoped; the protocol boundary decides what an agent can read.
The graph is kept in a structure built for fine-tuning models of your own, when you choose to. Your working data compounds into an asset you keep, not exhaust someone else mines.
For developers
Your fleet is a first-class citizen of your terminal. Dispatch a task from Claude Code, keep working, and the result comes back into your session as a proposal.
> /agent atlas draft follow-ups for the three quiet prospects
● fleet(dispatch atlas)
⎿ accepted · running async on atlas's node · reports back to this session
> tighten the retry backoff in src/retry.ts
● Update(src/retry.ts)
⎿ Updated src/retry.ts with 2 additions and 1 removal
● Bash(npm test)
⎿ 47 passing (2.8s)
● Backoff now doubles to a 30s cap. Tests stay green.
atlas · done in 4m · 3 drafts → proposal #159 in your queue
nothing sent yet. review with /queue, or from your phone.
- dispatch
One command hands a task to a named agent, from your editor or your terminal.
- async
Agents run on their own nodes while you keep your thread.
- governed
What comes back is a proposal in your queue, not an action already taken.
Private by design
A fleet this close to your operation runs where you decide, on accounts you own.
- self-hosted
Your infrastructure
Runs on your own hardware, a VPS you manage, or an EU cloud region. You choose where the fleet lives and who can reach it.
- your accounts
Your accounts, your model access
Agents reach your email, calendar, CRM, and books through your accounts, with credentials scoped per agent. Model access works the same way: your keys, your choice of provider.
- your data
Your data stays yours
Nothing your fleet does is used to build or improve anyone's models. The record of its work lands in your audit trail, on infrastructure you control.
Immutable audit trails.
#153inboxreply sent: invoice queryapproved by you17:21
#152outreachfollow-up sent: quiet prospectapproved by you15:47
#151financeinvoice issued and filedapproved by you12:03
#150finance2 transactions categorizedapproved by you10:26
#149researchweekly digest sentapproved by you09:14
Get your fleet
Havenware is in production and taking on new operations.
Talk to us
Evaluating Havenware for a company, a team, or an institution? Book a call. We’ll walk through your operation and scope a deployment.
Book a callOr write to hello@havenware.dev.