Anna
What this SDK contributes to the agent stack, and — just as importantly — what it does not.
Read this before judging what you see
src/anna/ is a contract: it declares the plugin, five tool schemas, two specialists and an autonomy guard, and it imports nothing and calls nothing. The reasoning loop — the part that decides which tool to call — lives in the anshin-anna service (GitLab 152), reached over HMAC-SHA256 at POST /api/orchestrate/decide, which is not running here. What is real below: the tool schemas, the autonomy guard, and the tools executing against the live wiki on db-01 — through anshin-wiki-server, which gates and audits every call.And one thing is deliberately less real than it looks. The autonomy guard below runs in your browser. It used to run in a Next.js API route, which made it look like a server-side control; that route is gone, and it was never the real enforcement point anyway — it was this demo enforcing a tier policy it had also invented. The guard is a pre-check. The control on this page is the permission gate, which is a different process in a different language, and the console reports the two separately rather than showing one number for both.
Plugin
wiki
v1.6.0
Tools
5
declared by the contract
Specialists
2
wiki · compliance
Plugin registration
- name
- wiki
- version
- 1.6.0
- product_id
- anshin-wiki-sdk
- description
- Wiki knowledge base tools and specialists for Anna AI
Specialists
Each specialist declares the subset of tools it may call.
wiki
General wiki knowledge base specialist. Answers questions about wiki documents, finds information, and helps with document management.
compliance
Compliance-focused specialist for HIPAA, SOC2, ISO27001, and NIST framework validation against wiki documents.
Tools
Risk level maps to a minimum autonomy tier through the SDK's own guard (getToolAutonomyLevel).
| search_wiki_documents | low | ADVISORY | Search wiki documents by keyword or phrase. Returns matching documents with relevance scores and highlighted snippets. |
| summarize_wiki_article | low | ADVISORY | Retrieve the content of a wiki document for summarization. Returns the document's text content that can be summarized by the LLM. |
| check_wiki_compliance | low | ADVISORY | Validate a wiki document against a compliance framework (HIPAA, SOC2, ISO27001, NIST). Returns the document content for LLM-based compliance analysis. |
| suggest_wiki_edits | medium | SUPERVISED | Retrieve a wiki document's content for edit suggestions. Returns the document text for LLM-based analysis of grammar, clarity, completeness, and compliance. |
| create_wiki_draft | medium | SUPERVISED | Create a new draft wiki document. Requires user confirmation before saving. |
Run a tool
The guard runs before any request is made. Drop the tier below a tool's required tier and the call is refused with the guard's own reason string — not a 500, and not a silent success. Two badges appear: guard is that client-side decision, and HTTP is what the server answered. Only the second one is a control. Two of the five tools have no local executor and make no request at all, which the output says in words rather than returning an empty result.