Peer CEO and CTO executives Β· 8 departments Β· 54 skill manuals.
Claude, Inc. organizes a virtual company around your project. The CEO turns your brief into workstreams, delegates to useful departments, collects files and review evidence, and resumes from a local project workspace next session. Your project can be a product, a business operation, a research effort, or something that spans several departments. You do not have to choose a template first.
The 54 employees are skill manuals, not 54 processes running in the background. Claude Code supplies the active assistant session, native agent tools, configured model, and permissions. The company supplies its roles, operating instructions, and a local task-and-decision record.
Explore the company Β· Project workspace guide Β· Architecture and alternatives Β· Roadmap
Claude Code plugin: install in Claude Code, then open your project folder:
/plugin marketplace add alebgl77/claude-inc
/plugin install claude-inc@claude-inc
/claude-inc:company Build the first version of my appointment-booking business for tutors
The CEO clarifies material unknowns and uses the packaged helper to initialize
or resume project state when available. It selects the departments your project
needs; all eight remain available.
Plugin commands use the claude-inc: namespace. Directly installed commands
use /company instead.
CLI and local files: install the company, including the company command:
curl -fsSL https://raw.githubusercontent.com/alebgl77/claude-inc/main/install.sh | bashOr inspect a clone and run its installer:
git clone https://github.com/alebgl77/claude-inc
cd claude-inc
./install.shOn Windows, run ./install.ps1 from the clone. The installer compiles
company.exe locally with the Windows .NET Framework compiler; it requires
Bash, such as Git for Windows. No launcher binary is downloaded. Python 3.9+
is required for project and mission commands; direct Python entry points also
work without Bash. See the
Windows instructions.
Use --brief-file founder-brief.txt for literal text on Windows PowerShell 5.1:
that shell can alter embedded quotes before a native command receives them.
PowerShell 7 preserves the tested literal arguments. The native launcher
replaces an unchanged, manifest-owned company.cmd transactionally; modified
or unowned launchers stop installation for inspection.
Use ./install.sh --project or ./install.ps1 -Project to install into the
current project's .claude/ directory. --no-bin / -NoBin installs only the
Claude Code files. -NoBin neither compiles nor migrates the Windows launcher;
it does not repair an existing company.cmd. Installers retain their collision checks and managed-file
transactions; they stop before replacing unmanaged or modified content.
A project install includes .claude/commands/company.md; invoke /company
explicitly when you want CEO coordination.
From the folder where the actual work belongs:
company project init --name "Tutor desk" \
--brief "Build a booking product for independent tutors, from offer to working prototype." \
--departments developers,designers,marketing \
--goal "A tutor can create a booking link and a student can request a lesson" \
--constraint "Use supplied facts; leave publication and purchases to me"
company project startinit creates .claude/company/project.json with your brief, goals, constraints,
and routing preferences. It does not fabricate tasks or alter your onboarding
profile. start launches a real Claude Code session in that project with the
packaged company plugin and resume instructions. It uses Claude Code's normal
permissions and configured model; it does not start a daemon or enforce a spend
budget. Claude Code must already be available and configured.
The CEO sets business priorities; the peer CTO sets technical direction across architecture, agent infrastructure, security, skills, and code. Their handoff is finding β business impact β options/tradeoffs β recommendation β decision needed, with technical acceptance criteria ready for the delivery team.
The CEO records task contracts, starts work after its dependencies are accepted, and delegates with the host's available agent tools. Departments produce actual files. Submission records the files' SHA-256 hashes and moves a task to review; an acceptance records a different department or CEO as reviewer and requires those files to be unchanged. Hashes establish file identity, not correctness; reviewer labels are declarations, not authentication. Review the evidence.
company project status
company project status --format json
company project start # resume from the same folder
company project prompt # inspect or copy the resume instructionsThe workspace guide covers task commands, blocked work, review, privacy, and using a different assistant. The structured project record remains the task source of truth. A Board Memo or Markdown ledger is a report derived from it.
For project work that needs explicit review gates, choose a stage and effort:
company project harness profiles
company project status --format json # read the current revision
company project harness generate --stage build --effort balanced --expected-revision 1
company project loop next --format json
company project loop promptReplace 1 with the current revision. Enabling a harness is an explicit choice:
schema-1 projects remain unchanged until activation. The CEO selects useful
skills and project-specific criteria; the local compiler checks that plan and
adds the task contract and two business checks. Light, balanced, and deep allow
2, 3, and 5 lifetime submissions per task. These limits count submitted attempts,
not model calls, time, or spending. A recorded extension can raise a cap to 10.
Review uses the current loop next template, a different reviewer label, actual
submitted evidence files, and passing results for every gate. The helper checks
the recorded contract and artifact hashes; it does not run tests or judge the
quality of the work. Existing accepted work stays historical. See the
complete harness and review example.
graph TD
subgraph EXEC["Peer executives"]
CEO["CEO / business direction"]
CTO["CTO / technical direction"]
CEO <--> CTO
end
CEO --> DEV["Developers / 6 skills"]
CEO --> DES["Designers / 6 skills"]
CEO --> MKT["Marketing / 6 skills"]
CEO --> SOC["Social Media / 6 skills"]
CEO --> FIN["Finance / 6 skills"]
CEO --> SMB["Small Business / 6 skills"]
CEO --> LEG["Legal / 6 skills"]
CEO --> SAL["Sales / 6 skills"]
CEO -.-> COS["chief-of-staff"]
FIN -.-> TOK["token-accountant"]
CTO -.-> TECH["cto-advisor / skill-vetting / appsec-review / agent-evaluation"]
The CEO owns business coordination and serializes project updates to prevent competing writes; that implementation rule does not subordinate the CTO. The CTO owns technical direction and fitness across the departments. Department agents apply the relevant employee manuals. Independent assignments can run in parallel when the host supports it; the company does not assume subagents can spawn nested subagents. A host without delegation can use explicit department passes, with that limitation made visible.
Onboarding chooses preferred departments and skills; it does not remove the rest of the company or create project tasks:
/claude-inc:onboard # plugin, project profile
/claude-inc:onboard --global # plugin, global profile
company onboard # CLI onboarding
company team # inspect active routing preferences
Installer users can opt in with --onboard or -Onboard. With --no-bin or
-NoBin, onboarding is deferred to the plugin. Project profiles take precedence
over global profiles. company brief and /company obtain only normalized
scope, departments, and skills from the profile validator, never its free-form
body or research status. An invalid project profile blocks global fallback.
Stored preferences do not authorize network access.
Both commands never read profile files directly.
The plugin form of the CEO command is /claude-inc:company; its onboarding and
standup commands are /claude-inc:onboard and /claude-inc:standup.
Profiles remain separate from .claude/company/project.json; project commands
leave company-team.md untouched. Saving profiles uses the existing strict
schema, 32,768-byte limit, safe temporary write, and atomic replacement. Replacing
an existing profile still requires the existing explicit replacement option.
Mission Studio provides five
starting recipes: launch, validate, release, proposal, and content.
They are convenient examples inside the broader company, not its boundary.
Open studio/missions.html locally or use the terminal:
company missions
company mission launch --brief "Launch my invoicing app for freelancers" --format promptRecipes compile a self-contained prompt with selected manuals, staged handoffs, and evidence checks. They do not initialize your project or execute its tasks. The browser's cards and preset links omit your edited brief. Manual byte counts compare source files; they are not token, cost, or runtime estimates. Read the Mission Studio guide.
For a focused request, use a department directly:
company dev "my tests fail after the last refactor"
company legal "review the NDA in nda.md"
company marketing "draft ad variants using the supplied product facts"Skills can also be invoked directly. The CEO is useful when work needs coordination, dependencies, decisions, or continuity across sessions.
For a focused technical decision, use company cto "review our integration plan" --print.
It includes the peer CTO charter and only its four staff manuals. CTO decisions
and CEO business arbitration should unblock teams without unnecessary status ceremony.
π¨βπ» Developers - VP of Engineering
| Employee | Role | Superpower |
|---|---|---|
superpowers |
Skill Forge | 14-skill power pack |
context7 |
Docs Fetcher | Live library docs |
mcp-builder |
Tool Wright | Wire up MCP servers |
skill-creator |
Skill Smith | Build your own skills |
webapp-testing |
QA Engineer | Browser-test your app |
claude-mem |
Memory Keeper | Persistent memory |
π¨ Designers - VP of Design
| Employee | Role | Superpower |
|---|---|---|
ui-ux-pro-max |
Design Lead | Full UI/UX system |
taste |
Taste Maker | Design-taste critic |
frontend-design |
Front of House | Build front-end UIs |
transitions |
Motion Artist | CSS motion library |
web-artifacts |
Prototyper | Live web prototypes |
brand-guidelines |
Brand Keeper | Build a brand kit |
π£ Marketing - CMO
| Employee | Role | Superpower |
|---|---|---|
copywriting |
Word Smith | High-converting copy |
ai-seo |
Search Whisperer | Rank in AI search |
cro |
Conversion Lead | Lift conversion rates |
ad-creative |
Ad Maker | Ad headlines & visuals |
customer-research |
Voice of Customer | Synthesise user voice |
lead-magnets |
Bait Master | Build lead magnets |
π± Social Media - Head of Social
| Employee | Role | Superpower |
|---|---|---|
post-writer |
Ghostwriter | Write LinkedIn posts |
profile-optimizer |
Profile Doctor | Optimise your profile |
reels-scripting |
Reel Writer | Script your Reels |
hook-generator |
Hook Smith | Scroll-stopping hooks |
voice-builder |
Voice Coach | Clone your voice |
youtube-thumbnail |
Cover Tester | Test thumbnail covers |
π° Finance - CFO
| Employee | Role | Superpower |
|---|---|---|
financial-statements |
Statement Builder | Build the statements |
journal-entry |
Journal Keeper | Post journal entries |
reconciliation |
Reconciler | Reconcile the books |
variance-analysis |
Variance Analyst | Explain the variances |
audit-support |
Auditor | Prep for audit |
close-management |
The Closer | Run the close |
πͺ Small Business - COO
| Employee | Role | Superpower |
|---|---|---|
cash-flow-snapshot |
Cash Watcher | Snapshot your cash |
invoice-chase |
Debt Chaser | Chase late invoices |
plan-payroll |
Payroll Planner | Plan payroll |
margin-analyzer |
Margin Analyst | Analyse margins |
tax-prep |
Tax Prepper | Prep your taxes |
run-campaign |
Campaign Runner | Run a campaign |
βοΈ Legal - General Counsel
| Employee | Role | Superpower |
|---|---|---|
review-contract |
Contract Reviewer | Review any contract |
triage-nda |
NDA Triage | Fast NDA review |
compliance-check |
Compliance Officer | Check compliance |
legal-risk-assessment |
Risk Assessor | Flag legal risk |
vendor-check |
Vendor Vetter | Vet a vendor |
signature-request |
Signature Wrangler | Route for signature |
π€ Sales - VP of Sales (Series A hire)
| Employee | Role | Superpower |
|---|---|---|
account-research |
Prospector | Actionable prospect intel |
draft-outreach |
Cold Emailer | Outreach that earns replies |
call-prep |
Deal Prepper | Walk into calls sharp |
proposal-builder |
Rainmaker | Proposals that close |
objection-handler |
Persuader | Turn pushback into progress |
pipeline-review |
Pipeline Doctor | Pipeline truth, weekly plan |
ποΈ Executive staff - attached to the C-suite (Series A hires)
| Employee | Reports to | Superpower |
|---|---|---|
chief-of-staff |
CEO | Project context, task contracts, decision log, weekly review |
token-accountant |
CFO | Reports from observed or supplied usage records |
cto-advisor |
CTO | Technical constraints, architecture decisions, skill selection |
skill-vetting |
CTO | Source/license/hash review, static risk, separate signature verification |
appsec-review |
CTO | Defensive threat models, diffs, dependency and secrets hygiene |
agent-evaluation |
CTO | Paired skill trials, measured results, uncertainty and costs |
The company combines native host roles with a small local state helper:
| On the org chart | In this repo | Mechanism |
|---|---|---|
| CEO | commands/company.md; /claude-inc:company in the plugin, /company in direct installs |
Business direction, priorities, arbitration, and serialized state updates |
| CTO | agents/cto.md; company cto |
Peer executive: technical direction, architecture, infrastructure, security, skills, and code |
| 8 departments | Eight department charters in agents/ |
Native host agents; independent assignments may run in parallel |
| 48 employees | skills/*/SKILL.md |
Skills with trigger-rich descriptions; VPs hire them per task, or they self-trigger |
| 6 staff skills | Two coordination/usage manuals and four CTO manuals | Skill packaging; this grouping does not define the CTO's executive authority |
| Project workspace | .claude/company/project.json |
Validated tasks, dependencies, artifact hashes, reviews, and decisions |
| Optional harness | Local schema-2 task policies and loop guidance | Required review gates, lifetime submission caps, and explicit extensions |
you β project brief β CEO β department assignments β files + review β project state + Board Memo
The roles and manuals available to your project. Open the full organization map.
Every employee follows the same contract: When to use β Workflow β Output format β Quality bar β Example. That's what makes all 54 manageable and PRs reviewable.
And the company audits itself: python3 scripts/validate.py (run in CI on every push) checks every job description, cross-references the CLI roster against the departments, and fails the build if an employee is hired twice, orphaned, or missing from the docs.
The company CLI composes self-contained prompts: a department charter plus its six manuals, or the peer CTO charter plus its four manuals. If claude is installed it runs it; otherwise pipe it anywhere:
company roster # meet the team
company team # show active routing preferences
company onboard --print # inspect the onboarding prompt
company brief "launch my product" # CEO mode
company finance "reconcile bank.csv vs ledger.csv"
company design "critique screenshot.png" --print | gemini # any engine
CLAUDE_INC_ENGINE=codex company dev "add tests" # or set an engineThe role manuals travel as plain text. company project start specifically
uses native Claude Code so its agents, skills, and plugin context are available.
For another assistant, inspect company project prompt and provide the named
local files and tools it needs; delegation, permissions, and persistence depend
on that host. A copied prompt does not turn every assistant into a compatible
execution adapter.
Onboarding can flag missing capabilities. It researches candidates only after separate, explicit consent and returns suggestions only. Each candidate is checked for source quality, current maintenance, compatibility, documentation, license, security, third-party adoption and available performance evidence. Stars include a collection date and never decide the ranking alone. Remote content is inspected read-only, treated as untrusted and never saved locally. Candidates fail if they are abandoned, opaque, unauditable, insecure, incompatible or poorly documented. During this onboarding research, no third-party content is downloaded, copied, installed or executed.
The four CTO manuals are first-party procedures. skill-vetting also includes
an unchanged, commit-pinned NVIDIA Skill Inspector reference
with its Apache-2.0 license and provenance.
It is reference material inside one skill, not a 55th registered employee.
The local wrapper's evidence format, permissions, and trust boundaries take precedence.
NVIDIA SkillSpector and SkillEvaluator, Cisco's scanner, and linked Trail of Bits
plugins remain optional external tools. They are not installed automatically;
Trail of Bits skill text is not bundled. Missing tools are NOT RUN. Source
review, runtime usefulness, and signature provenance are different checks;
none gives this company or a candidate NVIDIA certification. Source snapshots
were checked on 2026-09-08 and do not update themselves. The project's Python
runtime remains 3.9+; the optional SkillSpector 2.11.1 snapshot requires Python
3.12 through 3.14. See the vetting manual.
The proposed roadmap asks what it would take for one changed business decision to reach every affected department, with useful results the founder can inspect. Four horizons set the order:
- Next: prove a real company delivery through repeated local pilots and human review.
- After that: connect decisions across teams, preserving accepted history while proposing affected rework.
- Then: let teams earn their place through measured staffing and skill trials.
- Exploration: fork the company to compare isolated business scenarios before choosing a direction.
These are future proposals with evidence gates, not shipped capabilities or release dates. The founder keeps the final mandate; CEO and CTO bring the business and technical tradeoffs together.
Do I need a company for every request? No. Use a skill or department for a focused change. Use the CEO and project workspace when ownership, dependencies, reviews, and continuity help.
Does /company use model tokens? Yes, when the host executes work. Activating a department can create another model context. Manual-size comparisons do not predict the total usage, and Claude, Inc. does not enforce a budget.
Finance/Legal outputs? Decision support requiring appropriate professional review. The department manuals require limitations to be stated; instructions cannot guarantee an assistant follows them.
Can I hire more employees? Yes: we're hiring. One PR = one new employee.
- Org chart concept: Build Your Whole Team with Claude by Charlie Hills; this repo is the runnable formalisation of that map.
- Some employees are self-contained homages to great ecosystem projects: obra/superpowers, Context7, claude-mem.
- Built with Claude Code subagents, skills and plugins.
MIT for first-party content. The vendored NVIDIA reference retains its Apache-2.0 license.
If your new workforce ships something, β star the company - it's cheaper than payroll.
