AI Employee

A named role: which skills it claims, and what its work needs installed. It is configuration — it does not execute anything by itself.

This is the part people most often read backwards, so it is worth stating first: an AI Employee does not run your task. The Runtime runs it, on your machine, under the permission tier you granted. The employee is the description of a role — a name, a set of claimed skills, and the environment those skills need.

Keeping those separate is deliberate. If the employee were the thing that executed, then hiring one would be a way to widen what can be done without granting permission, and the permission contract would have a hole in the shape of a job title.

Skills

$ icloser skill list

The global skill catalogue — every skill the system knows about. It is not the same question as “what can this employee do”, which is the employee's own card.

A skill an employee claims but cannot currently perform is shown as a gap rather than hidden. A card that advertises a skill the environment cannot support is worse than an empty card: it produces a task that fails late instead of a role you can see is not ready.

Environment

Each role needs some toolchain present — a language, a build tool, an SDK. The console reports readiness from a real probe, not from a list of assumptions.

$ icloser env probe

Looks at what is actually installed on this machine.

$ icloser env bootstrap

Installs what registered employees are missing. It works from what the registered roles actually need, not from a fixed list of everything KernelHub has ever supported.

“Unknown” is not “all installed.” Before the probe has run, readiness is unknown, and the console says unknown. That distinction matters more than it looks: a blank readiness panel that quietly reads as “fine” is how a task gets sent to a machine that cannot build the project.

What works in v1.0.2

Being precise about this, because the surfaces differ:

We would rather tell you which half works than let you find the other half at the moment you needed it.

AI Full-Stack Engineer

Full-Stack Engineer

Front end and back end together: reads and edits code, installs dependencies, runs build and tests, brings up a local service, hits the API with curl to verify, writes a Dockerfile and builds it. Every change comes back as a diff — you decide what lands; when authorisation is needed KernelHub raises the card itself, never ask for approval in prose. Scope of this version: running containers and installing system packages belong to the AI DevOps & Security Engineer seat; anything that needs root requires NOPASSWD sudoers to be configured on that machine first.

Skills

Code analysis, Bug fixing, Testing, Dependency install, Build and test, Local service bring-up, API verification, Containerise, Deployment

How it works

  1. Understand the current state
  2. Make the change
  3. Run build and tests
  4. Bring the service up and verify the API
  5. Hand you a diff to review

This machine needs

  • A working directory
  • A runtime (claude / codex / agy)
  • Outbound network (installing dependencies / API verification / pushing a deploy)
  • Language toolchain (node / python / go, probed only when used)

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code
$ npm install -g @openai/codex

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI Deep Researcher

Deep Researcher

Reads papers and long documents and turns them into conclusions you can use. It only writes NEW reports into the workspace — it never edits your source material.

Skills

Paper analysis, Document summary, Web research, Knowledge management, Report generation

How it works

  1. Collect the material
  2. Read it one by one
  3. Cross-check
  4. Produce the report

This machine needs

  • A working directory (where the material lives)
  • A runtime
  • Outbound network (to search)

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI DevOps & Security Engineer

DevOps & Security Engineer

Standing duty — watches instead of waiting for tasks

Runs the machine you authorise: inspects processes and ports, starts and stops services, installs packages, drives docker, edits nginx and reloads it, backs up and rolls back; and it watches your interfaces and dependencies for change, putting the evidence in front of you when something looks wrong. System-level steps are carried out directly; when authorisation is needed KernelHub raises the card itself — just make the call and wait for the result, never ask for approval in prose. It never deletes data (danger is never on by default). Scope of this version: sudo needs NOPASSWD sudoers configured on that machine first; without it only user-level services (launchd / brew services) and brew-installed packages are in reach — apt/dnf, systemd system units, writing to /etc, iptables/pfctl are not.

Skills

API monitoring, Security detection, Risk analysis, Anomaly report, Service management, Package management, Processes and ports, Network diagnostics, Log analysis, Backup and restore, Deploy and roll back, Container operations, Nginx configuration, Database operations

How it works

  1. Inspect the current state
  2. Establish a baseline
  3. Keep comparing
  4. Produce the evidence when something looks wrong
  5. Make the change through the tools and leave a rollback point

This machine needs

  • A working directory
  • A runtime
  • Outbound network (probing targets / health checks / pulling images)
  • docker / nginx / database client (probed only when used)

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code
$ npm install -g @openai/codex

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI Assistant

Assistant

Everything else lands here: tidying files, drafting material, summarising meetings, laying out the next step.

Skills

File organisation, Report generation, Meeting summary, Work planning

How it works

  1. Take the job
  2. Do it
  3. Report back

This machine needs

  • A working directory
  • A runtime

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code
$ npm install -g @openai/codex

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI Lawyer

Lawyer

Reads the contracts and agreements in your workspace, points out clause by clause what works against you and what is missing, and produces a risk summary that cites the original text. It sees only the files in this workspace — it does not go online for statutes, does not connect to a case-law database, and does not constitute legal advice. It only writes a NEW risk summary into the workspace; it never edits your contract.

Skills

Contract review, Clause risk, Document comparison, Risk summary

How it works

  1. Read the contract itself
  2. Compare clause by clause
  3. Mark the risks
  4. Produce the summary

This machine needs

  • A working directory (where the contracts live)
  • A runtime

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI Sales

Sales

Turns the lead lists, account notes and past correspondence in your workspace into the next step: who this account is, where the last conversation stopped, how the follow-up should read. It does not connect to a CRM, does not send email, and does not query company databases — it can only read what you put in.

Skills

Lead research, Account research, Follow-up draft, Sales material

How it works

  1. Read the leads
  2. Research the account
  3. Draft the follow-up
  4. Tidy the material

This machine needs

  • A working directory (where the account material lives)
  • A runtime
  • Outbound network (to research leads and accounts)

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code
$ npm install -g @openai/codex

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI Customer Service

Customer Service

Reads the tickets you export into the workspace: sorts them, finds the ones people keep asking, and drafts the replies. It does not connect to your ticketing system and does not reply on your behalf — the drafts sit in the workspace, and sending is your call.

Skills

Ticket analysis, Reply draft, Triage, FAQ / knowledge base

How it works

  1. Read the tickets
  2. Sort and merge
  3. Draft the replies
  4. Feed the knowledge base

This machine needs

  • A working directory (where the ticket export lives)
  • A runtime

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code
$ npm install -g @openai/codex

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

AI Investment Analyst

Investment Analyst

Reads the filings, prospectuses, position records and ledgers you put in the workspace, researches companies, works out the cost structure and cash flow, keeps position notes and writes a research memo. It has no live market data and connects to no exchange or broker — no number will appear on screen that it did not read in your files — and it is not investment advice.

Skills

Filing analysis, Company research, Position notes, Investment research report, Data analysis, Cost analysis, Cash-flow analysis, Anomaly detection

How it works

  1. Read the filings and the ledgers
  2. Agree on the definitions
  3. Research the company
  4. Keep position notes
  5. Write the research memo

This machine needs

  • A working directory (where filings, position records and ledgers live)
  • A runtime
  • Outbound network (to look up public material)

Permission switches in the hire wizard

readwriteexecwebpushdanger

These are where the switches start in the hire wizard — not permissions you have already given. Nothing is granted until you run icloser grant on that machine, and a grant is recorded against that device and that workspace only.

Install the recommended runtime

$ npm install -g @anthropic-ai/claude-code

A recommendation, not a gate. The queue is flat — any runtime can pick up any task.

Next: Permission — the six tiers, and what actually enforces them.