The Forge Family: An Agent Operating System on Claude Code

Repeating the same multi-step professional work, in marketing or in research, usually means rewriting the same instructions, forgetting the checks, and losing the record of what was decided. I wanted reusable command systems, each with its own roles, review steps and files, that I could combine without rebuilding them.

There are four layers, all on Claude Code.

Command systems. The Forge Suite has 14 members for marketing and commerce, among them web, SEO, brand, social, ads and support systems, plus a credentials vault. They share a core: a media pipeline with a verify gate, a debate protocol, and governance made of an approval queue, a budget ledger, an audit log and a kill switch. The research suite has 18 skills, among them manuscript, case, grant and review systems, and 21 sub-agents such as a citation verifier and a fact-checker. I use it on my own manuscripts, cases and teaching.

Applications. GrowthForge (over the Forge Suite) and ScholarForge (over the research suite) are Tauri desktop apps with a Rust shell, a React and TypeScript front end and a Python sidecar. By design they never re-implement the systems and never call the Anthropic API directly. They run the local claude command headless and read the systems’ own files. GrowthForge has 12 screens and a sidecar self-test that passes 66 of 66. It is feature-complete for go-live and blocked only on code signing. ScholarForge has reached a read-only first milestone.

Operating product. Helm is the successor to the Forge Suite, and its plan carries all 183 of the old subcommands into its actions, roles and skills. Helm has its own page, including its failures.

Research instrument. Brand House is a Python standard-library harness with 40 offline tests. It runs an AI organisation through a debate of 22 turns with a three-proposer council and five enforced stages, then through downstream work. Every hand-off goes into a hash-chained append-only ledger, and roles reach files only through a logged MCP tool. I built it to study whether brand decisions actually govern the work an AI produces.

The evidence is about use and tests, not outcomes. The research suite is in daily use. GrowthForge passes its self-test and is waiting on code signing for distribution. Brand House passes its 40 offline tests.

I have no measurement showing that the marketing systems improved any business result, and I do not know of one in these files. The research data from Brand House is tied to a paper under submission, so I do not report it here.

Governance belongs in the architecture. The approval queue, budget ledger, audit log and kill switch were in the shared core from the start, so each member inherited them. That is cheaper than bolting controls onto each tool afterwards.

A desktop app over a command system should not duplicate it. Keeping the apps thin, so that the systems stay the source of truth, avoids two versions of the same logic drifting apart.

Finally, these are agent systems on a frontier model. Where a small local model has a place, I have kept it to low-risk lanes. That is why I describe small-model work separately, with its measurements.