Layer 3 — Arbol as an MCP server. A retrieval and reasoning layer over an Obsidian vault with knowledge graph RAG capabilities.
Arbol provides 8 well-named, read-only tools with clear descriptions (avg 180 chars, within baseline 194). All tools have input schemas with types and required fields. However, parameter descriptions are sparse or missing entirely (e.g., 'k', 'hops', 'budget_tokens' lack context). Output schemas are undocumented, LLMs cannot predict response structure. Error handling is minimal; tools return plain strings rather than structured errors with recovery guidance. No tool annotations (readOnlyHint, idempotentHint) despite all being read-only. Composition is strong: tools chain well (search → note → near), and naming follows verb_noun convention (vault_search, vault_context). Security is solid: no credentials exposed, all operations are read-only. Baseline comparison: avg tool description 194 chars (Arbol ~180 ✓), avg params 4 (Arbol 2 - 6 ✓), but 100% of A+ tools have param descriptions (Arbol ~40% ✗) and documented return types (Arbol 0% ✗).
Attention-worthy items: open questions, blocked work, items needing review, long-draft notes, orphans, and substantive notes misfiled as MOCs.
Assemble a token-budgeted, citation-manifested context pack for a question. Prefer this over vault_search when you are about to write a substantive answer: it returns the manifest that every claim must be traceable to, and warns when the index is stale.
List the notes linked to a given note, by hop distance. Use to understand what a note sits between.
Read one note in full by its exact title, with its frontmatter, backlinks and outlinks.
The vault's hierarchy from Home down. Read this first in a new session.
Search the Obsidian vault and return matching passages with their chunk_id, note title, section heading and file path. Use this before answering any question about the vault's projects, its own schema, or past decisions — the vault, not your priors, is the source of truth for those.
No output schemas documented. LLMs cannot predict response structure (e.g., vault_search returns markdown string, vault_stats returns JSON). Tools return unstructured text or raw JSON without field definitions.
Parameter descriptions missing or minimal. 'k' (appears in vault_search, vault_context) lacks explanation of what it controls (result count). 'hops' described only as 'graph expansion radius; 0 disables', no guidance on typical values or performance impact. 'budget_tokens' unexplained.
No tool annotations. All 8 tools are read-only and idempotent, but lack readOnlyHint and idempotentHint annotations. LLMs cannot infer safety properties and may incorrectly treat these as stateful operations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Corpus statistics: note and chunk counts, total words, types distribution.
Notes by last-updated, newest first. Optionally scoped to a project.
Error handling returns plain strings without recovery guidance. E.g., 'No matches. Try fewer or more distinctive terms.' is helpful but not structured. No error categorization (retryable vs user-fixable). Tools do not return actionable next steps for failures like 'No note titled X'.
vault_context returns a citation manifest as markdown string, not structured JSON. LLMs must parse text to extract chunk_ids and note references, increasing error likelihood and token waste.