ClawCompany: a 38-role AI company OS that runs from npx

What ClawCompany actually solves, and for whom

Most agent frameworks give you one assistant and a tool list. ClawCompany's premise is organisational instead: you act as Chairman, issue a mission, and a named set of roles decomposes and executes it. The README's worked example is a stock analysis, where a Leader splits the request into work streams, a Researcher gathers data through web search and a price feed, an Analyst builds a model with a code interpreter, and a Writer formats the output. The stated cost for that mission is $0.06.

The intended audience is people who are not developers. The README's comparison table puts ClawCompany under "Everyone", against OpenClaw for developers and Paperclip for technical users, and the setup claim is deliberately minimal: npx, one key, no Docker, no config files, no proxy. If you have ever wanted a research or analysis pipeline that produces a document rather than a chat transcript, that is the shape being sold here.

There is a second audience the README does not advertise: engineers who want to read a TypeScript implementation of role decomposition, tool dispatch and tiered memory. The repository is a pnpm workspace with server, cli, packages and templates directories, so the mechanism is inspectable rather than hidden behind a hosted API.

The mechanism: roles, tools, and a think-act-observe loop

Three pieces fit together. The first is the role roster. Six templates ship in the repository's templates directory: Default with 9 roles (CEO, CTO, CFO, CMO, Researcher, Analyst, Engineer and others), YC Startup with 7, Trading Desk with 7, Research Lab with 5, Software Dev with 6, and Harness Builder with 3. The counts sum to 37 in the table as published, while the README headline says 38 roles; the discrepancy is not explained in the README, and it is small enough that you should count the template files yourself if the exact roster matters to you.

The second piece is the tool layer. Nine tools are listed: Web Search, Web Fetch, Price Feed, Browser Use, Shell, Filesystem, HTTP, Code Interpreter and Memory Search. The README states that every role gets all nine, and that agents drive them through a think, act, observe loop. Uniform tool access is a real design decision, not a detail: it means a CFO role can run shell commands, so the trust boundary is your machine, not the role definition.

The third piece is model routing. Professional work is directed at Opus or Sonnet, routine work at a cheaper Flash-Lite class model, with the README claiming this is about 30 times cheaper than routing everything to Opus. The routing rule itself is not documented in the README, so treat the cost figures as the project's own claims rather than measured numbers.

The 4-layer memory and where the 400-token figure comes from

Memory is the part with the most specific claims attached. Layer 4 is the Chairman layer: your preferences, language and industry, always injected. Layer 3 is the company layer, split into four partitions the README names as Culture, Decisions, Learnings and Tech Stack, auto-categorised and compressed by an LLM. Layer 2 is an archive of compressed originals, kept searchable so that nothing is discarded. Layer 1 is the current session context.

The design intent is context economy. The README states that roughly 400 tokens are injected per mission, against 5,000 or more for OpenClaw, and characterises the result as the same quality at one thirteenth of the cost. That is a comparison from the project's own README with no published methodology, so the number to trust is the structural one: four tiers, one of them compressed, one of them an archive. Whether 400 tokens is enough context for your domain is something you can only judge by running a mission and reading what gets injected.

The archive layer is the more interesting choice. Keeping compressed originals and making them searchable through the Memory Search tool means the system is designed to be lossy at the injection layer but lossless at the storage layer. That only works if retrieval is good, and the README does not describe the retrieval method.

Installing ClawCompany and running a first mission

The quick start is a single command. The README states it takes two steps and about 30 seconds, and requires Node.js 20 or newer; the package.json engines field confirms node >=20.

npx clawcompany

For a source checkout, the README gives a pnpm flow. The repository pins pnpm@9.15.0 in package.json and uses a pnpm-workspace.yaml, so pnpm is the expected package manager rather than npm.

git clone https://github.com/Claw-Company/clawcompany.git
cd clawcompany
pnpm install
pnpm dev

The README states the API runs at http://localhost:3200, which matches the PORT=3200 value in .env.example. Configuration is copied from .env.example. The only required key is ClawAPI; Anthropic, OpenAI, Google and DeepSeek are listed as optional direct providers, and all of those lines are commented out in the example file.

cp .env.example .env
# then set CLAWAPI_KEY=sk-claw-your_key_here

Database configuration is the one place where the example file leaves a choice open. DATABASE_URL is commented out with the note that leaving it unset uses embedded PGlite. The root package.json exposes db:generate and db:migrate scripts that delegate to the @clawcompany/db package, so a first run against a real Postgres instance means setting DATABASE_URL and running the migration script before starting the server. What you should see after pnpm dev is the API listening on port 3200; the README does not describe the first screen of the UI, so do not expect a documented onboarding flow.

Limitations: unproven cadence, uniform tool access, thin docs

The clearest limitation is stated by the project itself. The README says ClawCompany "continues as a free open-source option" and that "maintenance cadence is slower as we focus on Sagit", the closed-source macOS sibling. The last push to main was on 2026-04-25. No releases were retrieved. If you need a dependency with a predictable upgrade path, this is not it, and the README is honest about why.

A second issue is the trust model. Every role gets all nine tools, including Shell and Filesystem, and agents act autonomously. There is no role-level permission scoping described in the README. A trading-desk template and a research-lab template therefore have the same reach into your machine. If you are evaluating this for anything touching production credentials, that is the first thing to check in the code.

Third, documentation depth. Several load-bearing claims are asserted without method: the $0.06 per mission figure, the 30x routing saving, and the 400-token injection. The role count in the headline does not match the template table. The README does not document rollback, does not describe how auto-categorisation decides between Culture, Decisions, Learnings and Tech Stack, and does not explain what happens when a mission fails midway. None of that makes the project wrong, but it means you are reading a product page as much as a manual.

Finally, the case where it is the wrong tool: if you want a single well-scoped agent with a narrow tool allowlist and a stable API surface, a smaller framework will be less to reason about. ClawCompany's value is the org chart, and you pay for it in moving parts.

ClawCompany compared with OpenClaw and Paperclip

The README's own table sets up the comparison, and the differences are architectural rather than cosmetic. OpenClaw is described as aimed at developers, with one agent, no templates, flat-file memory, and an npm plus config JSON setup costing $0.40 or more per mission. Paperclip is described as aimed at technical users, bringing your own roles, with no templates and no memory layer, installed through Docker and Postgres.

The distinction that matters is where the structure lives. OpenClaw leaves orchestration to you and stores memory as flat files. Paperclip leaves the role design to you and stores nothing. ClawCompany ships the roles and the templates as files in the repository, and treats memory as a four-tier subsystem with LLM compression and an archive. That is a different bet: less assembly, more opinion, and more code you inherit.

The setup difference follows from that bet. Docker plus Postgres, as in Paperclip, is heavier to start but easier to operate as a service. npx plus one key is lighter to start but leaves you running a local process and an embedded PGlite database by default, with the operational questions that come with it. The README does not compare the two on durability or backup, which is the gap a team evaluating this for real work would want filled.

Maintenance cost, licensing and what to verify before you commit

The licence is MIT, stated in the README and in package.json, with the copyright line reading MIT © 2026 ClawCompany. MIT permits commercial use, modification and redistribution with the licence text retained. The README's invitation to fork is consistent with that. This is a description of the licence terms, not legal advice; if you are embedding this in a product, read the LICENSE file in the repository root yourself.

The upgrade cost is the harder question. There are no retrieved releases, the root package.json is marked private at version 0.1.0, and the workspace pins pnpm@9.15.0 with TypeScript ^5.9.3 and vitest ^4.1.0 in devDependencies. A private root package with a 0.1.0 version and no releases means there is no published changelog to diff against, so tracking upstream changes means reading commits. Combined with the stated slower cadence, the practical posture is to treat your fork as the maintained artifact rather than the upstream repository.

Before you commit, verify the database path. .env.example leaves DATABASE_URL commented out and notes that leaving it unset uses embedded PGlite, while db:migrate exists as a script. Confirm which of the two you are actually running, because that decision determines where your company memory physically lives. The README's closing line, "Your AI company remembers everything, locally", is only true if you know which store is holding it.

Editorial conclusion

Adopt ClawCompany if you want to inspect how a multi-role agent company is wired together in TypeScript, or if you want a local, MIT-licensed memory layer you can fork. Do not adopt it if you need a vendor with a release history and a support commitment; the README states that maintenance cadence is slower as the team focuses on the closed-source Sagit desktop app, and the last push to main was on 2026-04-25. Before committing, verify three things yourself: that your preferred provider key works through CLAWAPI_KEY or one of the optional provider variables in .env.example, that pnpm db:migrate runs against your chosen database, and that the memory partitions the README names (Culture, Decisions, Learnings, Tech Stack) match what you want stored on your machine.

Frequently asked questions

Official sources

1. Claw-Company/clawcompany on GitHub
2. Issues
3. License: MIT
4. README