mcp-aurekai
Official<p align="center">
<img src="https://raw.githubusercontent.com/aurekai/aurekai/main/assets/aurekai-logo.svg" alt="Aurekai" width="520" />
</p>
# `@aurekai/mcp` — Aurekai MCP Server
**`0.8.0-alpha.5`** · capability-native · zero dependencies · stdio + Streamable HTTP
Exposes all 9 Aurekai capability families (111 commands) as MCP tools with full protocol-level features:
tool annotations, resource pagination, named prompts, `_meta` proof propagation, and embedded resource outputs.
## Install
```bash
npm install -g @aurekai/mcp
```
## Usage
### stdio (default — for Claude Desktop, Cursor, etc.)
```jsonc
// claude_desktop_config.json
{
"mcpServers": {
"aurekai": {
"command": "aurekai-mcp"
}
}
}
```
### Streamable HTTP (optional)
```bash
AKAI_MCP_HTTP_PORT=3100 aurekai-mcp
# POST JSON-RPC to http://127.0.0.1:3100/mcp
```
## Protocol Surface
| Feature | Status |
|---|---|
| `tools/list` — 89 operators across 9 capability families | ✅ |
| Tool annotations (`readOnlyHint`, `destructiveHint`, `idempotentHint`) | ✅ |
| `resources/list` — 13 `aurekai://` resource URIs | ✅ |
| `resources/read` — live reads for runtime/capabilities, queue/stats, models | ✅ |
| Resource pagination (`nextCursor`) | ✅ |
| Resource subscriptions (acknowledge) | ✅ |
| `prompts/list` + `prompts/get` — 8 named capability prompts | ✅ |
| `_meta` proof propagation on tool call results | ✅ |
| Embedded resource outputs for proof-emitting tools | ✅ |
| `logging` server capability | ✅ |
| Streamable HTTP transport (`AKAI_MCP_HTTP_PORT`) | ✅ |
## Capability Families
| Family | Operators | Examples |
|---|---|---|
| `runtime` | 11 | `akai_api`, `akai_queue`, `akai_workflow` |
| `commerce` | 11 | `akai_gate`, `akai_pay`, `akai_ledger` |
| `intake` | 12 | `akai_transcribe`, `akai_ingest`, `akai_segment` |
| `memory` | 11 | `akai_fpq`, `akai_fpqx`, `akai_embed`, `akai_vec` |
| `proof` | 8 | `akai_proof`, `akai_canon`, `akai_graph`, `akai_hash` |
| `reason` | 5 | `akai_reason`, `akai_physics`, `akai_flow`, `akai_learn` |
| `wire` | 5 | `akai_tel`, `akai_wire`, `akai_moq`, `akai_net` |
| `publish` | 9 | `akai_brief`, `akai_narrate`, `akai_pack`, `akai_distribute` |
| `substrate` | 17 | `akai_capability`, `akai_space`, `akai_compress` |
## Named Prompts
| Prompt | Description |
|---|---|
| `turn-this-call-into-a-deliverable` | audio → transcribe → brief → deliverable |
| `inspect-this-artifact-lineage` | Resolve full Merkle lineage for an artifact |
| `build-a-model-memory-pack` | FPQ compress + roundtrip + export memory pack |
| `compare-these-reasoning-branches` | Dual branch diff with recommendation |
| `generate-client-invoice-from-usage` | Metering records → invoice |
| `produce-wire-device-report` | PCAP → SIP event + device report |
| `run-a-release-gate` | proof validate + manifest verify + SLI auto-run |
| `make-a-client-brief-from-this-audio` | audio → transcript → structured client brief |
## Resources (`aurekai://` URIs)
`aurekai://runtime/capabilities` · `aurekai://queue/stats` · `aurekai://ledger/portfolio`
`aurekai://models` · `aurekai://model-memory` · `aurekai://features/{artifact}`
`aurekai://proof/{id}` · `aurekai://graph/{node}/lineage` · `aurekai://space/{name}`
`aurekai://wire/{capture_id}` · `aurekai://project/{id}` · `aurekai://invoice/{id}` · `aurekai://cms/{entry_id}`
## Runtime Requirement
Tools require the `akai` binary on `PATH` (from [aurekai/native-runtime](https://github.com/aurekai/native-runtime))
or set `AKAI_BIN=/path/to/akai`. Without it, tools return a clear error message — no crash.
## Registry Targets
- [Smithery](https://smithery.ai/server/io.github.aurekai/aurekai-mcp)
- [Glama](https://glama.ai)
- [Official MCP Registry](https://mcp.so)
- PulseMCP
TDQS
Scored across 89 tools
With 89 tools, many have overlapping purposes despite broad categories. Tools like akai_fpq, akai_fpqx, akai_quant, and akai_flash_qla all relate to quantization/inference, while many 'operator' tools (e.g., akai_discip, akai_fragment) have vague descriptions making them hard to distinguish. The brief and jargon-heavy descriptions further hinder clear differentiation.
All tools share the 'akai_' prefix and use snake_case, but the second part varies between single words (akai_cli, akai_gen) and multi-word phrases (akai_detect_objects, akai_frame_extract). Some are verbs (akai_compress), others nouns (akai_model), showing no consistent verb_noun pattern. This inconsistency makes the set moderately predictable.
89 tools is far above the typical well-scoped range. While the server covers a broad domain (AI/ML pipelines), the sheer number overwhelms an agent's ability to efficiently select the right tool. Many tools appear to be low-level or specialized, suggesting the surface could be consolidated.
Given the large number of tools, the server likely covers many aspects of its domain, but the vague descriptions make it difficult to assess whether CRUD or lifecycle operations are complete. There appear to be many specialized inference and pipeline tools, but gaps in common operations like error handling or logging are apparent.