Assistants MCP and Skills - VVG-web/aurora-studio GitHub Wiki
Wiki home · related: Asking the base · Trust · design: architecture §5
Aurora meets assistants (Claude Code, Cursor, OpenCode, any MCP client) in two ways: MCP gives them the base as a tool, and skills teach them the procedures.
scripts/aurora_mcp.py is a small stdio server (JSON-RPC 2.0, standard library only — no SDK, no Node). It lets an
assistant search the base, read cards and ask questions while it works on your task, instead of you carrying context
into the chat by hand.
| Tool | Does |
|---|---|
kb_search |
find cards by meaning and words — name, trust status, gist |
kb_card |
read a card whole |
kb_context |
assemble a context pack by topic (generate mode = trusted knowledge; while the base is young, bootstrap, marked drafts join the pack) |
kb_index |
the table of contents of the base: a line per card, grouped by section |
artifact_spec |
how to produce an artifact of this project: template, folder, prompt, the final-text boundary |
kb_ask |
ask the base: the project's model answers from the cards with links (slow — tens of seconds) |
It only reads. The assistant does not pass git-guard and does not take part in trust — the engine computes trust
from tasks, not the assistant. kb_ask does not write a conversation journal; it never changes files.
One server — one base. Customers must not mix: a server is started per project (--project <root>) and its name
carries the project's slug (aurora-<slug>), so aurora-alpha.kb_search is never confused with
aurora-beta.kb_search. A failed call returns isError: true with a plain-language reason (an unknown tool, a card that
does not exist, a mode outside the list), not a block of usage text.
kit:mcp (or python3 aurora.py mcp) checks the server and prints a ready connection block for every project on the
machine. The format is the same for all clients:
{"mcpServers": {
"aurora-alpha": {"command": "python3",
"args": ["<project>/.aurora/scripts/aurora_mcp.py", "--project", "<project>"]}}}MCP servers are configured once per kit and live in the kit's local/mcp.json (the panel's Kit setup); tokens
stay there, the panel only shows a mask. MCP is not configured in a project. Writing to Jira is a right of the
kit alone.
Selftest without an assistant: python3 .aurora/scripts/aurora_mcp.py --selftest.
Three skills ship with the kit (skills/):
| Skill | For |
|---|---|
aurora-vault |
operating the framework: building and maintaining the base, trust, Decision Records, reviewing and generating artifacts; the procedure references behind the "procedure" commands |
aurora-grill |
taking an idea apart along a decision tree: rounds of questions with recommendations until no silent assumptions are left. The planner runs it before an artifact is produced; it is available in chat as well |
aurora-dev |
developing the engine itself: running tests and scenarios, covering changes. Not for projects that merely use Aurora |
Skills are copies, not links into the repository: install them with kit:skills --apply into ~/.claude/skills/
(other harnesses get a link to that folder). aurora.py new does it for you and aurora.py update --apply refreshes
them. Installing a skill never leaves the agent without it: the new copy is made next to the old one and swapped in.
In a conversation, /aurora-vault <command> runs a procedure command (kb:ingest, kb:decide, make:create, …) —
the work that needs meaning rather than a script. Mentions in a task: @path (a project file or folder), /skill
(its method goes into the task), @server (an MCP server is connected at once).
This wiki is generated from the wiki/ folder of the repository — edit the pages there, changes made here are overwritten. · RU-Home