Skill Claude CodeCodex
Use when about to launch any GPU/CUDA/PyTorch/Ollama workload, when wrapping a program that uses an LLM or image generator, when handling GPU coordination on Mandragora, or when working with the bots (im-gen, llm-via-telegram, crush, MCP server). Loads the full Mandragora gpu-lock contract — non-blocking…
Skill Claude CodeCodex
Use when the user wants to pass current task context to another AI agent (Gemini, Qwen, etc.) so they can continue mid-thought. Writes a structured baton-pass to /.ai-shared/handoffs/. Triggered explicitly via /handoff [to].
Skill Claude CodeCodex
Use when adding, changing, looking up, or auditing keybindings/hotkeys/shortcuts across the mandragora repo. Loads the cross-program inventory and the user's existing binding "language" (modifier conventions, key families, shared prefixes) before proposing changes. Triggers on "bind X to Y", "what does do", "find a…
Skill Claude CodeCodex
Use when the user wants to refresh Open Notebook (nb.mvr.ac) with the current state of the mandragora-desktop obsidian vault. Runs the nb-vault-sync CLI, which idempotently diffs /Documents/mandragora-desktop-obsidian-vault/ against the vault notebook in Open Notebook and POST/DELETEs sources to match. Manual trigger…
Skill Claude CodeCodex
Use when /etc/nixos/mandragora has uncommitted changes that span multiple unrelated topics and the user wants each topic committed separately. Walks the diff, groups hunks by concern, and commits each group as its own commit.
Skill Claude CodeCodex
Use when the user wants to resume work from a handoff written by another agent. Reads the latest open handoff addressed to Claude from /.ai-shared/handoffs/ and continues from there. Triggered explicitly via /pickup.
MCP server Gemini CLI
One of 2 in settings.json
MCP server "local-ai" as configured in mvrozanti/mandragora. Launched with local-ai-mcp-server.
Skill Claude CodeCodex
Use when a request has more than one defensible reading and guessing wrong would waste real work — vague asks ("make it better", "fix the thing"), a fork with no obvious default, or a scope/blast-radius call that cannot be settled from the repo. Elicits intent through multiple-choice questions whose stems name…
Skill Claude CodeCodex
Use when a decision will be made by looking — UI/UX, layout, visual identity, information architecture, a redesign or migration of an existing surface, "which of these should it be", "make it look better". Ships several working directions as one published, fully interactive HTML artifact wired to real data, then…
Skill Claude CodeCodex
Pause and critically evaluate the current approach before continuing work. Use when a task is complex, consequential, ambiguous, unexpectedly difficult, accumulating complexity, producing repeated failures, or when new information may invalidate the current plan. Also use proactively at meaningful milestones, before…
Skill Claude CodeCodex
Explicit-only (/hallucinate) divergent ideation and "behind the veil" meta-analysis. Generates many out-there-but-plausible ideas and filters them to the few that are both novel and feasible; scores and ranks a list of existing ideas; and inspects the goal itself — spotting XY problems, reformulating the goal into…
Skill Claude CodeCodex
Toggle chat mode — replies become two people talking, not a report. Use when the user types /chat, /chat on, or /chat off. While on, prose to the user is stripped to the minimum that carries the thought; commits, code, comments and docs are unaffected. Survives the whole session and is never carried into a handoff.
Skill Claude CodeCodex
Explicit-only (/devil). For exactly ONE reply, become the technical devil's advocate for a project or part of one — attack its design and code hard and specifically, and pair every attack with a concrete better way that keeps observable behaviour ABSOLUTELY identical (refactors, single sources of truth, abstractions…
Skill Claude CodeCodex
Use when a project has grown hard to change — a small request turns into a sprawling plan, the same idea lives in several places that disagree, agents keep getting lost in or re-breaking one area, or the operator says "this is all so extensive", "minify this", "the code needs abstractions/enforcement", "clean this up…
Skill Claude CodeCodex
Explicit-only (/convince). Write the shortest self-contained argument that justifies a claim, decision or recommendation to a reader who was not there — one paragraph of short, complete sentences in plain words, every one of them literally true, built around the single reason that carries the conclusion. Humane in…
Skill Claude CodeCodex
Use when a decision will be made by understanding how something works rather than by looking at it — architecture, service boundaries, data model or ontology ("what is X", "where does the truth live", "how should A and B relate"), API/protocol contracts, consistency and failure semantics, pipeline and workflow design…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: