We connect the pastto the future

Every AI model your company pays for, and the agents built on them, in one place.

Connect Claude, ChatGPT and Grok with your own keys. Build agents, put a manager over them, and keep one record of every call, what it was for and what it cost.

Book a pilot call Watch a manager agent work
Overview
Conversations813+6% on last month
People using AI14 of 15one hasn't started
Spend$297.84+4% on last month
Personal data caught111 blocked

Usage over time

Conversations per day, by provider

ClaudeChatGPTGrok

What AI was used for

Sorted from each first prompt

Coding
41%
Writing
21%
Sales & marketing
14%
Customer support
11%
Data & analysis
7%
Translation
6%

Teams

This month against the budget each team was given

TeamMostly forTop modelSpendBudget
EngineeringCodingClaude Opus$193.00
88%
SalesSales & marketingGPT-5$15.79
53%
MarketingWritingGPT-5$14.17
over by $0.17
The Overview screen, showing the last 30 days at an example company of 15 people. Try the menu on the left.

PaFu sits underneath your company's AI.

Models come and go. The agents you build, the rules you set and the record of what happened stay in PaFu. Moving an agent to a different model is one setting.

  1. Govern

    Every call by a person or an agent is logged with its purpose and cost. Rules for personal data, budgets and model access apply to both.

    Overview · Activity · Policies
  2. Manage

    Manager agents take a task, write a plan, hand each step to one of their agents, and combine the results into one answer.

    Managers
  3. Build

    An agent is a job, a set of instructions and a model. Make as many as your teams need, without writing code.

    Agents
  4. Connect

    Your own API keys for Anthropic, OpenAI and xAI, plus any provider that speaks the OpenAI format.

    Claude · ChatGPT · Grok · Mistral · your own

Agents that work together, with a manager on top.

Each agent runs on the model that suits its job: Grok for quick research, ChatGPT for copy, Claude for the numbers. They don't need to know about each other.

A manager does. Give it a task and it writes a plan, sends each step to the right agent, lets them work at the same time, then writes one answer from what they found. Every step is logged with its model and cost.

Campaign managermanages Researcher, Writer and Analyst
Plan a spring campaign for our 5-seat Startup Pass bundle
PlanClaude Opus 5.5 · 3.1s
  1. ResearcherFind what competitors offer for team tickets, and two numbers we can quote.
  2. WriterWrite one short email and one LinkedIn post for the bundle.
  3. AnalystEstimate sign-ups as a range, and what drives them.
ResearcherGrok 4 · 2.4s

Competitors and numbers

Two competitors run spring bundles at 15 to 20% off. Neither offers team seats. Group buyers decide in the last three weeks before early-bird ends.

WriterGPT-5 · 3.0s

Email and LinkedIn post

Subject: Bring the whole team for one price. Our 5-seat bundle saves you 20% when you book before 14 March. One link, one payment, five tickets.

AnalystClaude Sonnet 5.5 · 2.7s

Sign-up forecast

120 to 160 bundles, so 600 to 800 seats, based on last spring. The biggest swing factor is how early the LinkedIn push starts.

Final answerClaude Opus 5.5 · 4.2s

Start the LinkedIn push this week, aimed at teams of five or more. Lead with the one thing competitors don't offer: team seats at one price. Expect 120 to 160 bundles. The email and post are ready to review.

5 steps on 3 providers · $0.21 · all logged in Activity
A manager run from the PaFu prototype. Each agent uses a different provider, and the manager never does the work itself.

People and agents, on the same record.

The Activity log lists every call across all your providers, whether a person made it or an agent did. Filter by person, agent, team, model or purpose, and open any row to read it.

If PaFu sorts something into the wrong purpose, change it. The fix teaches the next one.

WhenPersonPromptPurposeModelCost
Today 10:31ResearcherFind what competitors offer for team ticketsvia Campaign managerResearchGrok 4$0.03
Today 10:14Selin T.Suggest three slogans for our spring campaignSales & marketingGPT-5$0.02
Today 09:51Burak S.Why does this SQL join return duplicate rows?CodingClaude Opus$0.14
Today 09:30Ece M.Check the refund to IBAN [masked] for a cancelled bookingIBAN maskedCustomer supportGPT-5 mini$0.01
Yesterday 16:02Ozan B.Forecast November ticket sales from these weekly numbersData & analysisClaude Opus$0.11
Activity this week. The Researcher row came from the manager run above. Ece's IBAN was masked before it reached the model.

Or ask every model yourself.

Not every job needs an agent. Type once in Command and the prompt goes to every model your team is allowed to use. The answers arrive side by side as they're written.

Follow-up questions continue each model's own conversation. When you're done, the whole exchange is already in the log, with its cost.

Pick one of the examples to send it.

CommandSelin T., Marketing
Suggest three slogans for our spring ticket campaign

Claude2.1s

1. Your whole team, one pass. 2. Early beats loud. 3. Meet the room before it fills.

ChatGPT2.6s

1. Get in early. 2. One pass, every stage. 3. Where next year's deals start.

Grok1.7s

1. Skip the queue, not the party. 2. Early bird, sharper deals. 3. Be there first.

Logged as Sales & marketing $0.06
Logged this week1,284 prompts, $412.60
  • Selin T.Sales & marketing$0.06No issues
  • Emre K.Coding$0.31No issues
  • Ece M.Customer support$0.01IBAN masked

Rules that apply to people and agents alike.

PaFu reads each prompt before it's sent, whether a person typed it or an agent wrote it. If it finds an email address, a phone number, an IBAN, a card number, a national ID or an API key, it's masked or stopped.

  • Data rules for personal data, secrets and project code names
  • Budgets per team, with a warning at 80%
  • Model access per team, for example Legal on Claude only, for its agents too
Reply to jonas.l@example.com, his ticket code never arrived

Check before sending. This prompt contains an email address.

Mask and sendSend as isEdit
Personal data

Email, phone, IBAN, card numbers, national ID

Warn the person
Secrets

API keys and passwords

Block
Watch words

Project Atlas

Warn the person
What Kerem from Support sees when he pastes a customer's email into a prompt, and the rules behind it.

Where the conversation history comes from

AI providers don't hand out a list of everything your team has asked. So PaFu collects it in three ways, and we're upfront about what each one covers.

Through PaFu

Everything sent from Command, and from any internal tool you point at PaFu instead of the provider, is logged as it happens.

Enterprise exports

ChatGPT Enterprise and Claude Enterprise let admins export workspace conversations. PaFu imports those files today. Automatic sync comes with the pilot.

Personal accounts

Not covered. If someone uses a private ChatGPT account at home, PaFu doesn't see it, and we don't think it should without their consent.

Employees see a notice that their AI use is recorded. Managers see team totals. Message content is only visible to the roles you choose, and every time someone opens one, it's logged.

Why we're building this

I run sales, and this year our team used three different AI tools. Nobody chose that. It just happened, one subscription at a time.

When finance asked what we were getting for the money, I couldn't answer. I didn't know who used what, for which work, or whether anyone had pasted a customer's details into a chatbot. PaFu is the tool I wanted that day.

From the founder of PaFu

The plan

Oct – Dec 2026

Working prototype, used by our own team. Connections, agents and managers, Command, Activity, policies and history import.

Feb – Mar 2027

Pilots with two or three companies. Agents run on our servers with run budgets, single sign-on and live enterprise sync.

Summer 2027

Agents that use tools. Web search, CRM and email, with a person approving anything that can't be undone.

Nov 2027

Public launch at Slush. Visitors give a manager agent a task at our booth and watch its team work.

Bring one task you'd hand to an agent.

We're looking for two or three companies for the first pilot. Leave your email and we'll build that agent with you on the call, on your own models.