Provider-neutral LLM binding layer for the Aurekai capability runtime. Exposes 48 operators across 9 capability families to all major model providers via a unified canonical tool schema, MCP bridge, evals harness, and gateway router.
| Family | Tools | Description |
|---|---|---|
runtime |
8 | Dispatch, enqueue, gate, doctor, recipe runner |
commerce |
6 | Auth, invoicing, metering, ledger, feature gates |
intake |
4 | Transcription, ingest, transcript cleaning, frame extraction |
memory |
10 | Model pull, FPQ compress, vector search, SAE activate, embed, cache |
proof |
7 | Bundle, hash, lineage graph, export, canon parse, inspect |
reason |
3 | Session start, branch, diff |
wire |
3 | Tel listen, PCAP ingest, wire report |
publish |
4 | Brief, deliverable, distribute, narrate |
substrate |
3 | Space open/put, capability registry |
| Provider | Tools | MCP | Schema Format | Evals |
|---|---|---|---|---|
| OpenAI | 48 | ✓ | tools[].function |
✓ |
| Anthropic | 48 | ✓ | tools[].input_schema |
✓ |
| Gemini | 48 | via bridge | function_declarations[] |
✓ |
| Mistral | 48 | ✓ | tools[].function |
✓ |
| Groq | 48 | ✓ | tools[].function |
✓ |
| xAI | 48 | via aurekai-mcp | tools[].function |
✓ |
| Perplexity | 48 | via aurekai-mcp | functions[] |
✓ |
| Cohere | 48 | via aurekai-mcp | parameter_definitions |
✓ |
| Local (Ollama/llama.cpp) | 48 | ✓ | functions[] |
✓ |
npm install
npm run schemas:generate # regenerate all 9 provider schemas from canonical
npm run evals:run # 9 providers × 8 metrics dry-run matrix
npm run gateway:demo # capability-aware provider router
npm run openai:demo # OpenAI remote MCP demoCanonical contract: schemas/aurekai-tools.canonical.json — single source of truth.
Provider schemas are generated; do not edit them manually.
node scripts/generate-provider-schemas.mjs
# → schemas/{openai,anthropic,gemini,mistral,groq,xai,perplexity,cohere,local}-tools.jsonnode evals/runner.mjs
# → output/evals/evals-report.json (9 providers × 8 metrics)
# → output/evals/{provider}.akrun.json + .akproof.jsonMetrics: latency_ms, cost_usd, cache_hit_rate, proof_completeness, retrieval_precision, feature_drift, tool_call_success, schema_compliance.
The capability router in gateway/index.mjs selects the optimal provider for a request based on declared requirements (latency, cost, proof, retrieval, local):
import { gatewayRequest } from "@aurekai/llm-integrations/gateway";
const result = await gatewayRequest({ tool: "aurekai.memory.vec_search", query: "..." }, { prefer: "local" });| Provider | Demo |
|---|---|
| OpenAI | examples/openai-remote-mcp.md |
| Anthropic | examples/anthropic-claude-desktop-mcp.md |
| Gemini | examples/gemini-cache-bridge.md |
| Groq | examples/groq-low-latency.md |
| Perplexity | examples/perplexity-citations.md |
| Mistral | examples/mistral-managed-connector.md |
| xAI | examples/xai-function-adapter.md |
| Cohere | examples/cohere-rerank-pack.md |
| Local | examples/local-llama-stack.md |
providers/ 9 provider adapters (adapter.mjs, run.mjs, manifest.json, README.md)
gateway/ capability-aware provider router
schemas/ canonical tool contract + 9 generated provider schemas
evals/ cross-provider benchmark harness + matrix.json
examples/ 9 provider demos
docs/ architecture, provider matrix, roadmap
scripts/ schema generator
All 48 tools are available via @aurekai/mcp — add to any MCP client:
{
"mcpServers": {
"aurekai": { "command": "npx", "args": ["-y", "@aurekai/mcp"] }
}
}MIT