foundry

Foundry — a creator plugin for Hermes Agent

Five creator bots sharing one structured memory ledger: research,
draft, score, and price your work — and remember what worked. Built
for the Hermes Agent Plugin Hackathon (Nov 2026).

Start here: say "foundry start" in chat. A short interview sets
up your brand (only the first question is required). Answer one
question at a time — your brand is saved as soon as the first
answer lands, and you can resume anytime by saying "foundry start"
again. Your brand lives in the Brand tab on the dashboard,
where you can view and edit it any time. Drafting refuses the
seeded demo brand until you have your own.

How a cycle flows

Patternmaker interviews you → Identity Card (your brand).
Prospector finds sourced evidence → Evidence Card (at least 2
sources with links; unconfirmed claims are dropped).
Smith drafts from cited Evidence Cards only, in your
brand's voice → Draft Card.
Proofing Table — you tweak (one-line note) or reject (reason);
feedback becomes a Revision Card and Smith returns a new
version. Approval is disabled, so nothing posts from this flow.
Inspector and Paymaster record performance and money for
the same draft; the next cycle learns from them.

Demo content

The ledger ships with demo cards — titled Demo: …, origin seed
— so the board isn't empty on first open. They're sample data: a
fictional brand ("three small tools"), illustrative numbers
(performance is labelled estimated). Never your voice, never your
numbers. The draft guard refuses the demo brand, and the agent
treats demo cards as samples, not sources. On a fresh install they
are the only cards until you run foundry_start.

The crew

Patternmaker — interviews you, owns the brand.
Prospector — sourced research (agent-supplied; no live web-search
surface exists in Hermes v0.23.0).
Smith — drafts from cited Evidence Cards only. A claim with no
source is cut. Live revisions use the host model via ctx.llm.
Inspector — performance: what to repeat, what to drop.
Paymaster — costs, revenue, profit, brand partnerships, tax
tracking (tracking only — no advice, no filing).

Every bot reads and writes the same versioned ledger cards. Feedback
becomes a Revision Card. Performance and money write back, so the
next cycle is smarter.

The board

Four tabs, on the web dashboard and in the Desktop app (opt-in):
Proofing Table (all drafts, newest first — Tweak / Reject feed
back to Smith), Brand, Inspector, Paymaster.

Approval is disabled, not implemented. There is no trusted human
review channel yet, so the Approve buttons are disabled and the
approve routes fail closed. Nothing auto-posts, auto-sends, or
auto-buys. Ever.

Tools (8)

foundry_start, foundry_ledger_list, foundry_ledger_get,
foundry_draft_create, foundry_draft_feedback,
foundry_draft_approve, foundry_scorecard, foundry_money_summary.

Install

cp -r foundry ~/.hermes/plugins/foundry
hermes plugins validate ~/.hermes/plugins/foundry --install-deps
hermes plugins doctor foundry --ci
hermes plugins enable foundry

Copy the inner foundry folder, not the whole project, or validation
fails with "no plugin.yaml". Validated against Hermes v0.23.0:
validate 16/16, doctor 8 tools, smoke test 34/34. Full record in
docs/VALIDATION.md.

Disclosures

Network calls: none. The plugin makes no calls to third-party
services. Smith's live revision path uses the host model's
ctx.llm (no external API key held by the plugin).
Inspector/Paymaster numbers in the seeded demo are illustrative
example data, labelled estimated on the dashboard.
Reads outside its own data: none. The plugin reads only its own
ledger, stored at plugin_db("foundry").
Shell commands / background processes: none.
Stored credentials: none. No API keys, no .env, no OAuth.
Telemetry / usage reporting: none.
Approvals: disabled (see above). Tweak/Reject feedback is
recorded as unverified, never as human authorship.
No self-updating code, no binaries, no .pyc in the tree.
Dependencies: none (Python stdlib only).

License

README from the reviewed commit 134692f; it updates when the author re-pins.

About