Coding
Fix issues, open pull requests, review code and triage bugs.
Default: Asks before any merge or deploy.
Meet Fumu in 60 seconds. Watch the intro
Message it from Telegram or the console. It plans the work, hands each step to a specialist agent, and asks you before anything risky. New skills plug straight in.
Why Fumu
Most agents are clever for a minute and forgetful forever. Fumu is built the other way round: durable, careful and accountable first, then clever.
A task waiting three days for a reply or an approval picks up exactly where it stopped.
Reads freely. Changes need your approval. Spending money and messaging as you always do.
Every answer comes with a full record of the steps it took and the approvals you gave.
Remembers your preferences, people and past decisions, and forgets them when you ask.
Human in the loop
Every action passes one checkpoint. When a step is above the trust you have given that capability, the task pauses and you get a card with the exact change: the code, the email, the amount.
Approve it, edit it or reject it. The task waits, for free, until you do. It survives restarts, reminds you once and then cancels safely.
Security and approvalsHow a request runs
Fumu understands what you want, writes a plan, and hands each step to the agent that knows that domain.
01
A message arrives from Telegram, the web console or a schedule you set. It becomes a task, linked to the conversation it came from.
02
Fumu works out what kind of request it is: a quick answer, a routine it already knows, or something new that needs a plan.
03
It brings in what it remembers about you, then writes a short plan, with each step handed to the right specialist.
04
Each specialist works through its step and checks the result. Every action is checked against your rules first.
05
A step above the trust you have given pauses the task and sends you an approval card. It waits until you answer.
06
Fumu checks the result against what you asked for, tries again if a step failed, and remembers what it learned.
07
The answer comes back where you asked, with a full record of what was done and what you approved.
Capabilities
Everything Fumu can do in the outside world is a capability you can add, pause or remove. Each new one arrives with no access, and follows the same approvals and the same record as everything else.
Single things it can do, each with a clear risk level.
Open a pull request · Find a meeting slot
Specialists that each know one area of your work.
Coding · Inbox · Research · Support
Multi-step jobs Fumu can run again, learned from what worked.
Weekly bill check · Prepare for a meeting
Capability packs
The first packs cover code, research, content, customer email and support. Each one is generic, so it fits any business, not just ours.
Fix issues, open pull requests, review code and triage bugs.
Default: Asks before any merge or deploy.
Track competitors, summarise trends and check pricing pages.
Default: None needed: the pack only reads.
Draft LinkedIn posts, articles and changelog announcements.
Default: Always asks before publishing.
Draft replies, onboarding emails and follow-ups from your own inbox.
Default: Asks before sending, until you decide it has earned your trust.
Triage support email, answer from your docs and escalate the rest.
Default: Asks before sending; labelling is allowed.
Autonomy
Leave the Inbox agent drafting for months while Research, which only reads, runs freely. Raise a level when a capability has earned it.
Level 0
Read and summarise only.
"What's in my inbox?"
Level 1
Draft, then ask before changing anything.
Drafts a reply and waits for your approval.
Level 2
Make changes that can be undone, tell you afterwards, and allow undo for a while.
Archives newsletters, accepts a meeting.
Level 3
Act without asking, inside limits you set.
Reorders groceries under €40.
How to read Fumu
Fumu changes face only when a task changes state. Glance at it and you know whether it is working, waiting, or waiting for you.
A new conversation, a first hello, nothing pending.
A message arrives from Telegram, the console or a schedule.
Working out what you asked and writing the plan.
Reporting back, with a full record of what it did.
A step finished, or you approved a draft.
Something big landed: a pull request with green tests. Rare on purpose.
It needs your approval, or something new contradicts what it believed.
Suggesting a rule after you approve the same thing again and again.
Waiting on your reply, a budget used up, or the kill switch pulled.
Where you talk to it
Wherever you message it, you get the same Fumu: the same memory, the same approval cards and the same record of what it did.
Chat, voice notes and approval buttons, from your phone.
Your tasks, approvals and the full record of what was done.
"Every weekday at 8, brief me." Fumu can start work on its own timetable.
For teams who already work in Slack.
Questions
Not without asking. Spending above a budget and messaging someone new as you always need your approval, whatever autonomy you have set.
That is normal. A task waiting three days for a reply or for your approval picks up exactly where it stopped, even if something restarts in between.
Yes. Every reply links to a full record of the steps it took, the actions it ran and the approvals you gave.
Yes. Ask "what do you know about my car?", then "forget that", and it is gone.
Journal

An agent that can send email as you is only useful if you trust it. Here is how Fumu earns that trust one capability at a time, instead of asking for all of it on day one.

Real errands involve waiting: for a reply, an approval or a Monday morning. Fumu is built so that waiting never loses work, and a retry never does anything twice.

Everything Fumu can do is a capability you can add, pause or remove. New ones arrive with no access, and earn trust the same way as everything else.
Watch the 60-second intro, follow a real run from issue to pull request, and ask for a live walkthrough.
See the showcase