The assistants

Four assistants that work on the platform — not on your customers.

Every vendor gives you a bot that answers customers. Mihu gives you four more that are pointed at the platform itself: one you run by asking, one that writes and deploys real code, one that tests agents before a customer ever meets them, and one that keeps making them better after launch.

Ask · Build · Test · Improve Real compute, not mockups 100% of conversations scored
01

Mihu Assistant

Talk to your platform.

A conversational layer over the whole platform. Create and update agents, connect a channel, launch a campaign, pull last week's numbers — by asking, not by clicking through six settings screens.

  • Create and update agents by describing the change you want
  • Connect channels, launch campaigns and pull reports from chat
  • Drive the same actions from Claude or ChatGPT over MCP at mcp.mihu.ai
  • It manages the features you already have — it does not write new ones
app.mihu.ai
02

Builder Assistant

Have it build you software.

This is the coding layer, and it writes real software. A custom JSON payload to your webhook. Contact segmentation with triggered flows. A connector to the legacy system nobody wants to touch. Then it deploys the result as a running service with its own compute — not a mockup, real infrastructure you can watch consume CPU.

  • Custom integrations the platform doesn't ship out of the box
  • A dedicated dev server with full file and SSH access
  • Run shell commands and read live logs
  • Automatic QA — code review, tests and spend tracking before it ships
builder · deploy #114

Deployed: crm-payload-mapper — a webhook transformer with a custom JSON schema.

vCPU0.5
Memory512 MB
Cost / mo$4.20
review passed qa 12/12 eu-west · 99.98%
03

Simulator Studio

Break it before your customers do.

It plays your customers back at you. Simulator Studio generates personas from your own conversations — the impatient caller, the code-switcher, the one leaning on your refund policy — runs your agent against them, and scores it on a skill scorecard. You find out where it breaks in a test run, not on a live call.

  • Personas built from your own conversations, never generic scripts
  • A skill scorecard on every run — resolution, empathy, policy adherence
  • Re-run the whole suite before any change ships
simulator-studio
Simulator Studio running an agent against generated customer personas
Resolution93
Policy adherence100
Empathy88
04

Improvement Agent

It gets better while you sleep.

The Quality & Coaching Agent reads every conversation — all of them, not the 2% a supervisor had time for — scores them against your rubric, finds what the agent didn't know, and coaches it. Ship on Monday and the agent is measurably better by Friday, without anyone opening a report.

  • 100% of interactions scored, human and AI alike
  • Knowledge gaps ranked by how often they actually cost you
  • Coaching written back into the agent automatically
  • The loop closes: what it learns feeds the next design
quality-coaching · last 30 days auto-coached
Resolution rate 7993+14
Empathy 8894+6
Policy adherence 96100+4
3 knowledge gaps closed 42 calls coached

Which one do I use?

Two of them talk. Two of them work.

The distinction people get wrong: Mihu Assistant manages what already exists. Builder Assistant writes what doesn't. Ask the first for a campaign; ask the second for a connector.

Assistant Use it when you want to… It does not…
Mihu Assistant Create an agent, connect WhatsApp, launch a campaign, pull last week's numbers — by asking. Write new software. It drives the features that already exist.
Builder Assistant Send a custom JSON payload to a webhook, segment contacts with triggered flows, reach a legacy system. Replace your day-to-day admin. It builds; it isn't the control panel.
Simulator Studio Know how the agent handles an angry caller, a code-switcher or an edge case — before go-live. Touch production. Every run is a simulation against generated personas.
Improvement Agent Score every conversation, find what the agent didn't know, and coach it automatically. Wait for you to read a dashboard. It writes the fix back itself.

See it run

Watch all four in one session.

Describe an agent, test it against a thousand simulated conversations, ship it to voice and WhatsApp, and watch it coach itself — on your channels, in your languages, wired to your CRM.