ClawCompany puts you in the chairman's chair and splits the work by role

You are the chairman, and the leader role does the decomposing

The framing is the product, and it is stated without irony in the section explaining what the thing is: you are the Chairman and your AI team executes autonomously. The worked example is a single line of user input, an instruction to analyse a ticker for a year, followed by the decomposition. A leader role turns it into work streams. A researcher gathers data using web search and a price feed. An analyst builds the model through a code interpreter. A writer formats the report. You then read the report, and the example gives a cost. The claim attached to that flow is about routing rather than about roles: professional work gets the expensive frontier models and routine work gets a lighter one, described as thirty times cheaper than running everything on the largest model. Two things are worth holding onto from this section. The output is a document, not a change to a system, so the failure mode is a bad report rather than a bad action. And the cost figure is an example, not a measurement, since nothing on the page says what the instruction was or what the models were.

The headline says 38 roles and the template table adds up to 37

There are six templates and each lists its own roles, and the arithmetic does not land on the number in the headline. The default template lists nine roles including chief executive, technology, finance and marketing officers, a researcher, an analyst and an engineer. The startup template lists seven around a founder coach, a product manager and a growth role. The trading desk lists seven around a fund manager, bull and bear analysts and a risk manager. The research lab lists five: a principal researcher, an experimenter and an evaluator. The software development template lists six, and the harness builder lists three in a generative-adversarial arrangement of planner, generator and evaluator. That is nine plus seven plus seven plus five plus six plus three, which is thirty-seven, while the title of the section and the project description both say thirty-eight. One role is unaccounted for, or one template's count is stale. It is a small discrepancy, and it matters more than it looks: a role is a prompt with a model assignment behind it, so if you are building on the template list you should diff it against the actual role definitions in the repository rather than trust the table.

Four memory layers, and roughly 400 tokens per mission

The memory design is the part that claims to be different, so the four layers are worth reading as a hierarchy. The chairman layer holds your preferences, language and industry, and is described as always injected. The company layer has four partitions, for culture, decisions, learnings and tech stack, which are auto-categorized and compressed by a model. The archive layer keeps compressed originals in a searchable form, with the promise that nothing is ever lost. The session layer is the current conversation. The cost claim attached to this is specific: about 400 tokens are injected per mission, against more than 5,000 for the tool being compared, at what the page calls the same quality and one thirteenth the cost. Two things to notice about the design. Compression happens at write time, not at read time, which is what makes the injected figure small and what makes the archive layer necessary to compensate. And the chairman layer being always injected means your standing preferences ride along on every mission, so a mistake in that file is a permanent tax on every run.

Nine tools, every role gets all of them, and the loop is think, act, observe

The tool list is nine names and there is no per-role restriction, which is stated explicitly: every role gets all nine tools. They are web search, web fetch, a price feed, browser use, shell, filesystem, HTTP, a code interpreter and memory search. Agents run them in a think, act, observe loop. The breadth is the point and the risk. A shell tool and a browser tool in the same list means a role described as a financial analyst can do considerably more than read a price feed, and the routing of work to roles is advisory rather than enforced by capability. The features listed further down the page lean into the same breadth: a code manager offering multiple terminals in one dashboard, a price feed for crypto and equities described as carrying no AI cost, an export path that turns mission reports into presentations, word documents or PDFs, chat channels for a web interface, Telegram and Discord with Slack described as coming, and five built-in providers with the option to add your own. The extraction targets include slides and documents, which is a reminder that whatever an agent can read, it can also be steered by.

The comparison table prices a mission at six cents against forty

There is a three-column table comparing this project with two others, and it is the most useful and least verifiable part of the page. On who each is for, the alternatives are developers and technical users while this one claims everyone. On roles, one alternative ships a single agent, the other expects you to bring your own, and this one claims thirty-eight built in. On templates, both alternatives have none. On memory, one uses flat files, the other has none, and this one claims a four-layer scheme with model compression. On setup, the alternatives want npm plus a config file or Docker plus Postgres, while this one claims a single command and one key. And on cost per mission, the first is given as forty cents and more, this one as six cents, with the middle one varying. Read as marketing, that is fair enough. Read as data, none of it is sourced: there is no benchmark description, no task, no model list, and no way to reproduce a forty-cent figure from the page. The architectural differences in the table are checkable in the repositories; the numbers are not.

The desktop sibling is closed source and the cadence has slowed because of it

The most honest paragraph in the README is at the top, and it is about maintenance rather than features. A sibling project called Sagit is described as the same AI company architecture packaged as a signed and notarized macOS application, with a polished interface, the same four-layer memory, a knowledge wiki and document delivery, built on what the page calls the same thin harness plus fat skill philosophy, with both projects bringing your own key. It is explicitly closed source. Then the sentence that matters for anyone who picked the repository to build on: ClawCompany continues as a free open-source option, maintenance cadence is slower as the effort focuses on Sagit, and contributions and forks are welcome. That is a candid statement of a project in maintenance mode with a commercial sibling on the same architecture, which is a perfectly normal arrangement and not a red flag. It does mean the open repository is where fixes land late, and the version in the package manifest is 0.1.0 with the default branch last pushed on 25 April 2026.

One required key, four optional ones, and an embedded database by default

The environment template is the clearest statement of the dependency surface. One variable is required: a key for the project's own model gateway, with the comment noting that it works for everything out of the box. Four more are present and commented out, for Anthropic, OpenAI, Google and DeepSeek, described as optional and there to use models from other providers. Then the server section, which holds a port number and a commented database URL with the instruction to leave it unset for an embedded Postgres-compatible engine. That last note is the interesting one, because it explains the claim of no configuration files: the default deployment does not need a database server because the data layer runs in process, and a real database is something you opt into. The install path matches the claim, a single command with no Docker and no proxy:

npx clawcompany

The README describes it as two steps and thirty seconds, and states Node.js 20 or newer as the requirement along with the five supported suppliers. For working on it rather than running it, the clone is followed by a pnpm install and a pnpm dev, with the API then served on port 3200. For development the workspace is a pnpm monorepo with per-package build and typecheck scripts, a test runner, a lint script, database generation and migration scripts, and the API served on port 3200.

Editorial conclusion

Adopt ClawCompany if you want a plausible multi-role decomposition of an instruction rather than one agent doing everything, and if the point of the exercise is the report at the end rather than an autonomous agent doing work you will audit. The templates are the feature to evaluate, since a trading desk and a research lab are different role sets for the same machinery. Do not adopt it as production infrastructure: the package is at version 0.1.0, the project states that maintenance cadence is slower because the effort went to a closed-source sibling, and the only published dependency is an MCP SDK. Four things to check first. Whether the role set in the template you want is real, since the table sums to fewer roles than the headline claims. Which model supplier you use, because cost claims are tied to routing routine work to a cheaper model family. Whether the memory layer behaves when you actually need to delete something. And whether you are comfortable that a closed-source desktop app is the actively developed surface.

Frequently asked questions

Official sources

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