All plugins

KyaniteLabs

GitHub profile7 plugins in the catalog285

Plugins by KyaniteLabs

7 plugins by KyaniteLabs in the catalog, most-starred first.

KyaniteLabskinocutGuardrailed local video editing: Kinocut's MCP server plus two workflow skills (editing and short-form repurposing). Trim, caption, reformat and check quality with local FFmpeg. Portable Agent Plugins v1 configuration and skills, no plugin code. Requires uv and FFmpeg on PATH; launches uvx kinocut==1.16.4 over stdio. First launch downloads Kinocut and dependencies from PyPI. URL inputs download the requested media; font and model-weight features fetch assets from GitHub. Explicit Revideo dependency installation runs npm ci and may download renderer dependencies. Local FFmpeg editing and color grading need no account or API key. Kinocut's optional vision calls default off and require explicit configuration: legacy product descriptions use an Anthropic-compatible endpoint, while sampled-frame QC uses Anthropic. The README explains how to use the host's image model instead, including local models. Revideo and Hyperframes telemetry default off. Unverified Hyperframes browser previews and Docker renders require explicit telemetry consent. Hyperframes capture/snapshots disable implicit model analysis. These controls are not a network sandbox; see the plugin README for configuration and remaining network use. Disclosure — launches uvx kinocut==1.16.4 over stdio, which downloads Kinocut and its dependencies from PyPI on first launch; edits and writes local video files with FFmpeg; URL inputs download the requested media, font and model-weight features fetch assets from GitHub, and an explicit Revideo install runs npm ci; optional vision features (off by default) send selected images or sampled frames to an Anthropic-compatible endpoint or Anthropic; Revideo and Hyperframes telemetry stay off unless the operator enables them.Communitytools271KyaniteLabsepochCalibrated time estimates on the Hermes Kanban board: when a worker claims a task, Epoch posts a Monte Carlo P50/P90 (plus P80/P95) estimate as a task comment, and when the run completes it records that run's wall-clock hours back to Epoch's ledger (auto_wallclock provenance, sanity-gated) so later estimates calibrate. A Hermes Desktop page charts a card's Monte Carlo spread three ways (distribution, chance-done-by curve, 50 dots); it is read-only and never writes the ledger. Dispatcher and worker profiles can differ; the estimate-to-run binding lives next to the ledger, Kanban comments are display only. Also adds an epoch_estimate tool and /epoch. Disclosure — runs npx -y @kyanitelabs/epoch@0.8.0 as a child process (downloads Epoch and its caret-ranged npm dependencies on first run, which the catalog pin does not cover; needs Node >=22), passing the task title and its three-point hint to the local CLI; the child env is scrubbed (allow-listed PATH/HOME/locale/Node/npm/proxy vars only, the parent environment is never passed through); estimates, actuals and run bindings live under <HERMES_HOME>/plugin-data/epoch/ and are shared by every profile; Epoch telemetry is forced off for plugin runs; no network besides the npm download, no API key; hook work runs on short-lived background threads and unfinished jobs resume when the plugin next loads; the Desktop page reads every Kanban board read-only through Hermes' internal kanban modules and runs a forecast in a throwaway data directory when you open a card; /epoch passes other Epoch subcommands through (it refuses serve, self-improve and telemetry/data mutations, and refuses ledger writes on gateway platforms unless opted in).Communitytools8KyaniteLabsag4teGit archaeology in Hermes Desktop: dig a local repository's history with DevArch and read it as strata, with commits per day stacked by kind of work, eras, quiet gaps, signals and turning points, weekly rhythm, lines added and removed, practice health, change size, hotspots, releases, and authors merged across spellings and split into people, agents and bots. Adds a Desktop page, /ag4te (list, show, status, dig, stop, resume) and a read-only ag4te_dig tool for agents. Disclosure — runs DevArch as a child process via uvx --from git+https://github.com/KyaniteLabs/ag4te@<pinned sha> devarch (uv downloads that pinned DevArch commit and its Python dependencies on first run, which the catalog pin does not cover; needs uv on the gateway host); reads git log of the one repository you choose, which must sit under the configured repo_roots (default ~/workspaces); stores digs under <HERMES_HOME>/plugin-data/ag4te; no LLM calls, no API key, no network besides the uv download; one dig at a time, each step time-limited (timeout_seconds); /ag4te dig and resume are off on chat platforms unless allow_dig_on_gateway is set.Communitytools3KyaniteLabsinnerscapeGentle check-ins, journaling and an Unstuck Coach with your own Innerscape server: save a mood check-in in your words, save a journal entry or park an inbox item only after you say yes to the exact text, and look back on today. When you're frozen or overloaded, the bundled unstuck skill names what's in the way, gives one next move and holds the rest. The skills never score, label or diagnose a mood. Adds four tools, /innerscape, two skills, a Desktop page and a dashboard tab. Disclosure — calls your Innerscape REST API (default http://127.0.0.1:3001) with INNERSCAPE_TOKEN, read through Hermes secret scope, which can read and write your whole account; plain http only to this machine (https to any server you configure), no proxies or redirects; tool results go to your model provider and stay in session history, and mood and journal text is sensitive, so your provider sees it like any other chat; consent is model-mediated: save tools refuse unless the assistant has asked you and set user_confirmed, and Hermes has no separate approval step (see README); the Desktop page reads today and saves a check-in or journal entry with that same token through the plugin's dashboard routes (/api/plugins/innerscape/), so anyone signed in to your Hermes dashboard can read and write your Innerscape account; reads and saves are off on chat platforms by default and saves are off in scheduled runs; no telemetry, no background jobs.Communitytools2KyaniteLabsseleniteTakes friction off the person, treated as accessibility: one test, "does this step, wait or check earn its cost to you?", applied to steps handed to a person, agents parking Kanban work on each other for reasons that aren't real, and checks that can't change the decision. Calm mode cues the agent toward no questions and one-line replies when your messages show overload. A read-only cleanse ranks what to clear across skills, plugins, hooks, approvals, runbooks and boards. Walls always stand: credentials, spend, publishing, destructive steps, legal, named freezes. Adds /frictionless, the selenite_ledger and selenite_show tools, two fake-gate tools, skills, a Desktop page and a dashboard tab. Disclosure — no network, no credentials and no shell from the plugin itself; /frictionless starts an agent turn through the gateway only when Hermes's allow_gateway_injection is on, otherwise it prints the text for you to send; selenite_show and ledger, cleanse and evolve tool calls send local Desktop events (short keys only, no paths or private text) that the open Selenite page reacts to; hooks add a context note in Calm mode, and in fake_gate_mode "nudge" (default) or "ask" they add a note to, or hand back once with a question, a kanban_block that reads as agent bureaucracy (they never approve, unblock, close or edit a card, and any error fails open); when you approve a safe command once or for the session, its approval key is logged as an allowlist candidate (denials, timeouts and dangerous commands are never logged); writes a metadata-only ledger (no block reasons or prompt text, 0600, rotates at 4 MB), per-card question counts and a counts-only snapshot under plugin-data/antibeaurocracy/, and reports under the Selenite state directory (SELENITE_STATE, else XDG state, else HERMES_HOME/plugin-data/selenite, else ~/.local/state/selenite); reads kanban.db read-only (the CLI audit's --all-homes flag adds every ~/.hermes* home); the cleanse pass (the selenite_ledger tool, the Desktop page and the dashboard tab can run it) reads, read-only, the hook commands in ~/.claude/settings.json, the skill folders of Claude Code, Codex, OpenCode and ~/.agents, and Hermes plugin manifests; the hermes antibeaurocracy audit CLI also reads agent rule files (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.gemini/GEMINI.md) and the SOUL.md/AGENTS.md of every Hermes home and profile; the bundled frictionless skill's miner script, when the agent runs it, reads your own messages from local agent transcripts (Claude Code, Codex and Hermes state.db, all read-only) and keeps only counts plus a few redacted samples (emails, IPs, tokens, secrets and your home path stripped), 0600, in that state directory; the CLI's opt-in --judge runs a command you name, the plugin never does; all hooks turn off with hooks: false or SELENITE_HOOKS=0.Communityautomation1KyaniteLabsiolitePick an LLM that is smart enough, then cheapest, in one tool call. Set an intelligence floor or name an anchor model, and Iolite returns the few models on the cost x intelligence front, ranked toward cheaper, smarter or balanced. It answers from a catalog of 621 models bundled in the plugin, so a fresh install works offline with no setup. Adds three tools (iolite_shortlist, iolite_model, iolite_show), /iolite, a skill and a Hermes Desktop pane that plots the catalog in 3D and 2D and follows the agent's picks. Data and credits: Epoch AI ECI (CC BY), LMArena ratings (CC BY 4.0), LiteLLM prices (MIT); every answer names its sources, and catalog rows whose source names Artificial Analysis are refused. Disclosure — no network unless you set catalog_url (one https GET, no credentials added; plain http only with allow_http); a post_api_request hook records tokens per second and per-task tokens and model seconds (task ids hashed) on your machine only, never uploaded, off with measure_speed and measure_usage false; your last shortlist, your own prices and temporary .pane-*.receipt markers are kept in the plugin data directory; three local Desktop routes (/landscape, /my-prices, /pane-receipt) add no auth of their own and rely on the dashboard's loopback bind, so keep the dashboard on 127.0.0.1; no prompt text is sent to the pane; no secrets read; tool results stay in session history like any tool output.CommunitytoolsKyaniteLabsunbl0ckOne queue of everything blocked on you: live agent prompts, pairing requests, blocked Kanban cards and chats on any platform that ask you something. Answer Kanban cards in place; optional one-line summaries and Ask send item text to your configured model through ctx.llm (both can be turned off). Disclosure — reads every local profile's chat history (state.db, read-only), profile titles and pending pairing requests (through Hermes's PairingStore, never the codes) to build one queue, records live clarify/approval prompt text in its plugin data, and can comment on and unblock Kanban cards when you answer one; by default sends each new queue item's text (other profiles' chats included) to the configured model for a one-line summary, and card data on Ask (turn off with bluf_enabled: false / ask_enabled: false); on messaging platforms /unbl0ck replies with a count only; uses Hermes's internal Kanban modules, so a Hermes update can break it; no network of its own, no shell, no credentials.Communitytools