w/06 · ai workspace · nihmbus.ai

Nihmbus

the command center for human + ai teams.

what

a command center for human + ai teams

codebase

147k lines · 229 commits

platform

web · rest api · mcp

stack

next.js 16 · postgres + pgvector · ai sdk 6

walk through the meadow

overview

Most teams already work with AI. It just lives in a dozen tabs — a chat here, an agent there, a scheduler somewhere else — and none of them know what the others did. Nihmbus is one workspace where the people and the AI teammates share the same missions, inbox, docs and decisions.

AI teammates take on real work — drafting, researching, replying, running browser sessions — at whatever level of autonomy the team sets. Anything that needs a human lands in a single decision queue, and everything either side does is on the record.

360+

actions in one registry

doors: ui, api & ai tool

workspace modules

nihmbus.ai — run your whole team. human and ai.

the challenge

An agent that can do anything is only useful if the team can see what it's doing, steer it, and stop it. Most AI tools pick one: a powerful agent nobody can supervise, or a safe assistant that can't actually touch the work.

Nihmbus had to give AI teammates real reach — every feature in the product — while keeping people in charge of the calls that matter. And it had to feel like a calm place to work, not a control panel.

01 · the team

people and agents, one roster.

AI teammates sit next to people in the same workspace: they get assigned missions, pick up inbox threads, draft content and leave notes in docs. They don't take seats — pricing is per human.

Each piece of work carries an autonomy level — supervised, semi-autonomous or fully autonomous — so a team can let an agent run the routine work and keep a hand on the rest.

→one decision queue for every approval
→an audit trail on every action, human or ai
→agents that work across all modules, not one silo

02 · missions

point the team at an outcome.

A mission is a goal with owners, a timeline and an autonomy setting. AI teammates break it into work, report back, and ask when they hit a decision they're not allowed to make on their own.

03 · one action, three doors

define it once. use it everywhere.

Every capability in Nihmbus is a single action: a Zod schema, a handler and a flag for whether it changes anything. From that one definition the product gets a UI action, an endpoint at /api/invoke and a tool the agent can call — so the agent can never do less than a person can, or more than the rules allow.

The same registry is served over MCP, so outside agents — Claude, Cursor, anything that speaks the protocol — can work in a Nihmbus workspace with scoped API keys.

→360+ actions, one dispatcher
→role gates and rate limits at the door
→mutating actions route through approvals

One action, three surfaces: UI, API and AI tool

04 · inbox

every conversation, one queue.

Slack, Discord and Telegram messages, social mentions and texts arrive in a unified inbox, matched to contacts and routed to whoever — or whatever — should answer. Agents draft the reply; people approve, edit or take over.

→contacts & crm with a dm router
→engagement tracking across channels
→sms and voice via twilio

05 · content

drafts with a reason attached.

The content engine proposes posts against the team's strategy and scores every candidate from 0 to 1 — pillar fit, novelty, voice, platform constraints and how similar posts performed — with the rationale written next to the number.

Novelty is measured in pgvector: a draft too close to something already written gets flagged before anyone wastes time on it.

06 · research & records

memory the whole team can search.

Research, notes and records go into one knowledge base, searched with a hybrid of 1536-dimension embeddings, trigram and full-text ranking fused together — so the agent finds the right document whether you remember the words or only the idea.

When the answer is on the web, AI teammates open real browser sessions through Browserbase and Stagehand, and bring back what they found as records.

07 · projects & docs

where the work actually lives.

Kanban boards that sync with GitHub Projects, collaborative docs with live cursors, design canvases, visual workflows with approval gates, and standups you can just talk through — realtime voice, transcribed into the record.

→real-time docs on tiptap + yjs
→workflows as graphs, with human gates
→github projects two-way sync

08 · extensions

teach it anything.

Extensions add skills — plain SKILL.md playbooks, compatible with ClawHub and Anthropic's agent skills — remote MCP tool servers, UI panels, named actions and their own workspace data. Permissions are declared and enforced, and secrets are sealed with AES-256-GCM.

09 · pricing

ai teammates don't take seats.

Starter is free. Team is $49 per human seat a month. Enterprise is a conversation. However many agents a team runs, the bill is for the people.

architecture

one registry, many doors.

A single Next.js 16 app serves the site, the workspace, the API and the MCP server, deployed on Vercel. A collaboration server for live docs runs alongside it on Railway, and Inngest handles the background work.

01
registry
360+ actions — zod schema + handler + mutating flag, defined once
→
02
surfaces
ui actions · /api/invoke dispatcher · agent tools · mcp server with scoped keys
→
03
agent
ai sdk 6 through hypery — chat, embeddings, voice, images — autonomy levels & approvals
→
04
data
postgres via drizzle · pgvector hnsw + trigram + full-text, fused with rrf
→
05
reach
browserbase + stagehand sessions · composio integrations · twilio · github
→
06
extensions
skill.md playbooks · remote mcp tools · module-federated ui panels
→
07
realtime
hocuspocus + yjs for docs · inngest for jobs and schedules

the developer experience

one definition. three doors.

from the Nihmbus platform page — the same handler runs when you click “New Mission”, call the API, or ask the agent.

missions.create

1# Create a mission via the API — same handler as the UI and the agent
2curl -X POST https://app.nihmbus.ai/api/invoke/missions.create \
3  -H "Authorization: Bearer $NIHMBUS_API_KEY" \
4  -H "Content-Type: application/json" \
5  -d {
6    "subjectId": "sub_abc123",
7    "name": "Q3 product launch",
8    "autonomyLevel": "supervised"
9  }
10 
11# The agent calls the same action as a tool:
12#   missions_create({ subjectId: "sub_abc123", name: "Q3 product launch" })

brand & design system

A black sheep with a teal face, its wool drawn like a cloud. The favicon outlines it in a gradient that never sits still — cyan through violet to amber — and the workspace stays quiet so the work, and the agents doing it, stand out.

teal

#00d1d1

shade

#00aeae

wool

#1f1f1f

ink

#141414

violet

#9b4fc4

amber

#f2a014

Geistheadlines & interface

Geist Monolabels, ids & code

·tokens in oklch, dark-first with a light theme
·the favicon's gradient animates — the sheep is always moving
·one type family, geist, from marketing site to workspace

by the numbers

229

commits

147k

lines of typescript

360+

actions

workspace modules

api routes

database migrations

counted from the source repositories

what we did

development

action registry & invoke dispatcher
workspace modules & data model
mcp server & extension runtime
realtime docs & github sync

design

workspace ux
decision queue & approvals
marketing site

brand

the sheep & its gradient
nihmbus.ai

ai teammates & autonomy levels
content scoring
hybrid search
browser sessions

nihmbus.ai ↗how it works ↗pricing ↗

next world

Hyperythe identity and payments layer for ai products. →