Loading ToolBench
MCP server for Akashic Aurora - a knowledge recall and task coordination system that learns from agent actions and surfaces relevant lessons at decision points
This server exhibits severe definition quality gaps across nearly all 18 tools. Most tool descriptions are cryptic, domain-specific jargon-heavy, and fail to explain WHAT the tool does or WHEN to use it in plain language. Parameter schemas are present but descriptions are minimal or absent. Output schemas are undocumented. Tools like 'knowledge_recall', 'recall_at', 'knowledge_full', 'knowledge_map', 'delta', 'sweep', and 'glance' use opaque terminology ('lessons', 'boot', 'ledger', 'bus flow', 'WorldSnapshot') without explaining what these concepts mean to an LLM. Error handling is absent. The server appears designed for a specific internal knowledge system (Akashic Aurora) rather than as a general-purpose MCP tool suite. LLMs would struggle to determine when and how to invoke most tools.
What moved since your last boot mark: commits, task-ledger transitions, and bus flow -- the full 'what changed' report your truncated boot block points at. Read-only (never advances the mark).
Find files by glob-like pattern under the project root. '*' spans directories. Vendored/cache dirs excluded.
Show a git diff (read-only).
Recent git commit history (read-only).
Show a commit's message and diff (read-only).
Short git status of the working tree (read-only).
T079/T060 bounded WorldSnapshot program projection over the git-durable task ledger. Returns source-derived identities, per-row epistemic provenance, loud UNCHECKABLE organs, and optionally a compact operational summary.
Opaque domain terminology without explanation. Tools like 'knowledge_recall', 'recall_at', 'knowledge_map', 'delta', 'sweep', 'glance' use internal jargon ('lessons', 'boot', 'ledger', 'bus flow', 'WorldSnapshot', 'durable movement') without defining what these mean. LLMs cannot infer domain semantics and will misuse these tools.
Descriptions under 20 characters or missing context. 'glance' has no description at all. 'delta' and 'sweep' descriptions are cryptic ('What moved since your last boot mark', 'Pure structured awareness'). These fail to explain WHAT the tool does or WHEN to call it.
Output schemas completely undocumented. No tool documents what fields it returns or what structure the LLM should expect. This forces LLMs to guess at response structure and breaks tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Assemble the project's startup context for a task (recent notes + top lessons), the same briefing an agent gets.
Pull the FULL stored record of ONE recalled lesson by its source pointer (e.g. 'learn:experiment:NAME'), including fields beyond the selected recommendation.
Walk the knowledge graph OUTWARD from a topic -- connected lessons, notes, and docs. knowledge_recall SEARCHES ('everything about X'); this BROWSES ('what is X connected to?'), surfacing lessons you didn't know to search for.
Search Akashic Aurora's learned-knowledge base (lessons/decisions) via the project's own recall door.
List files and subdirectories of a directory in the project.
PRIVATE note-to-self (your scratchpad, NOT shared project knowledge): a durable working note injected into YOUR future boots. Re-noting the same title supersedes the old note. Use for how-YOU-work notes, e.g. 'when reviewing T039, start at packet_spec.py'.
Read YOUR private scratchpad (notes-to-self from prior invocations; also auto-injected into your boot). Yours alone -- peers have their own.
Read a text file from the project. Prefer start_line/end_line for big files to save tokens.
Re-run recall-at-action with a higher limit to see MORE lessons that cleared the relevance floor. The one-hop pull when a tool result's recall block says 'N of M shown'. Same engine as the automatic injection, more entries.
Search file contents for a regular expression (grep). Returns path:line: text, capped.
Pure structured awareness for exactly one subject: bus depth/window, bench, routing attendance, and durable movement. Reads concurrently; advances no cursor and registers no presence.
No error handling guidance. Tools provide no recovery hints, error classification, or actionable error messages. An LLM receiving a failure has no guidance on whether to retry, ask the user, or abandon the task.
Parameter descriptions are minimal or absent. Many parameters lack explanation of what they control or what values are valid. E.g., 'novelty' in knowledge_recall is described as 'If true, tag each result [boot]/[new]', unclear what [boot] and [new] mean or why an LLM would choose true vs false.