GitHub profilePlugins by pgreyy
2 plugins by pgreyy in the catalog, most-starred first.
pgreyyattestPaper trading ledger for agents. Orders, balances and P&L live in a local SQLite ledger the plugin owns, and the agent gets its figures from get_portfolio, get_pnl and get_quote instead of stating them. Orders are priced from live public exchange quotes with slippage and a fee, after a per-order size limit, a daily loss kill switch and duplicate-order checks. No real money, no exchange account, no keys. Disclosure: network use is read-only HTTPS GETs of public ticker data, each sending only the trading pair and the User-Agent attest-paper-trading/0.6.1, with a 3 second deadline: www.okx.com /api/v5/market/ticker?instId=BTC-USDT first, then api.kraken.com /0/public/Ticker?pair=XBTUSDT, then data-api.binance.vision /api/v3/ticker/bookTicker?symbol=BTCUSDT as fallbacks. Requests happen on an order, on get_quote, when marking open positions, and for the claim check. It writes one SQLite ledger in its Hermes plugin data folder: account, trades (with the quote that priced each one, the full tool result and the optional order note), positions, risk state, each day's starting equity, and the ids of sessions that called an attest tool or the guarded tool (removed after 90 days); no other conversation text. pre_tool_call blocks only the one tool named in the intercept_tool setting (empty by default) while paper_mode is on, and simulates that order on the ledger instead; every other tool returns untouched. transform_llm_output appends one line under a reply that states a balance or P&L figure the ledger contradicts, only in sessions that used attest, and never edits the reply. /attest check runs the bundled self test in a separate Python process against a temporary database. The demo_stub_tool setting (off by default) registers a test-only demo_broker_order tool.CommunitytoolspgreyycustomsPlants fake credential files on your machine and blocks a hijacked agent the moment a tool call reaches for one, before the tool runs. It also redacts its own fake values from tool results and from the request to the model provider, reads local scripts before they run, and asks for approval when an upload follows a credential search and packaging. Disclosure — plants fake credential files (mode 600, never over an existing file, inside a Git repository only at paths Git ignores) under the home folder, including ~/.ssh and ~/.aws, when the user runs /customs plant, blend or add, and keeps its own state in its plugin data folder; it only recognizes its own bait files and fake values and does not find or protect real credentials. Its hooks read every tool call's arguments and results and every outgoing model request, block or send to Hermes approval the calls that touch bait or carry a fake value, redact fake values from tool results and model requests, and after a block require approval for network actions in that session until /customs unlock (on by default). It reads the local scripts a command is about to run (up to 8 files, 512 KB each) and, during /customs blend, folder names under the home folder, and runs git check-ignore when planting inside a Git repository. No network calls by default; with alert_target set, alerts go out through hermes send on the user's own Hermes messaging route and carry the rule, the bait file's path, the likely source and a masked preview of the call (first 200 characters).Communitytools
