- Home
- Skills
- Agents & Orchestration
- Hermes Decision Brain
Works with the AI Agents you already use
Hermes Decision Brain
A local-first decision router that cascades from CPU inference to cloud only when confidence is low.
$19
Hermes Decision Brain
Example session with this skill installed
**Route this user request:** "I want to cancel my subscription immediately because of a duplicate charge last month." **Choices:** cancel, upgrade, support, billing. Use auto mode.
- Read your context and instructions
- Compiled the hermes decision brain
Decision successfully resolved using the local tier. The state fit the local token budget and the model answered with enough confidence, so no cloud escalation was needed. ## Decision Audit | Attribute | Value | | :--- | :--- | | **Result** | `cancel` | | **Source** | `laya` (Local, free) | | **Confidence** | `0.7402` | | **Escalated** | `False` | | **Latency** | `211 ms` | | **Cost (USD)** | `$0.00000` | | **Model** | `laya-onnx-v1` | ```json { "decision": "cancel", "type": "choice", "confidence": 0.7402, "escalated": false, "escalate_reason": null, "cost_usd": 0.0, "latency_ms": 211, "source": "laya", "layer_detail": { "laya": { "status": "success", "confidence": 0.7402, "distribution": [0.7402, 0.0891, 0.1103, 0.0604], "state_tokens": 68 } }, "metadata": { "mode": "auto", "token_counter": "laya_tokenizer" } }
Connects securely to your tools. The creator never sees your data.
About this skill
Stop wasting budget on cloud LLMs for simple classification tasks. This skill provides a tiered decision engine that handles routine logic locally on your CPU and only escalates to the cloud when confidence drops below measured thresholds. ## What it does - **Multi-tier routing** handles decisions through a local ONNX model first, then an optimized cloud tier. - **Typed outputs** returns specific `noul` probabilities, `choice` selections, or `score` values rather than unpredictable prose. - **Automated escalation** triggers cloud calls based on state length, low confidence, or flat probability distributions. - **Audit tracking** provides detailed latency, cost, and confidence metrics for every decision layer. - **Runtime management** builds and maintains an isolated Python 3.11 environment for local inference. ## How it works 1. **Initialize the runtime** using the bundled installer and health check scripts to set up the local ONNX environment. 2. **Configure routing** by choosing between `auto`, `local`, `cloud`, or `strict` modes to control escalation behavior. 3. **Execute decisions** via `decision_router.decide()` by passing state, questions, and desired output types. 4. **Monitor performance** through the `on_escalate` callback to track costs and refine local confidence thresholds. ## Frameworks & tools Built for Python 3.8+ (router) and Python 3.11 (local venv). Uses ONNX for local CPU inference via Laya and integrates with OpenRouter for cloud escalation. ## Why this beats prompting it yourself Prompting an LLM for classification is slow and expensive. This skill replaces vague prompts with typed primitives and a local-first cascade that cuts latency to milliseconds for routine tasks. ## Use cases - **Intent classification** to route incoming user requests to specific agent tools. - **Sentiment analysis** for high-volume ticket triaging without cloud costs. - **Binary validation** to confirm if a state meets specific criteria before proceeding. - **Risk scoring** for fraud or churn detection using consistent numerical ranges. ## Known limitations Calibrated on five reference cases. Local performance on non-English languages is inconsistent. Requires manual setup of ONNX checkpoints as they are not bundled.
Thresholds need recalibration for production traffic. Local tier has a 512-token sequence window. Requires Laya ONNX checkpoints on disk. French `noul` accuracy is question-dependent: the reference churn-risk case measured 0.0259 (FR) vs 0.8211 (EN), while a refund case measured 0.9188 (FR, correct). See references/laya-checkpoints.md for the measured record.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 month ago
Requires
Permissions
Allowed Hosts
File Scopes
Network access: - openrouter.ai — only on cloud escalation (mode="auto" or "cloud"); never called in mode="local". - huggingface.co + pypi.org — only on explicit update_laya.py (user-initiated model/runtime update). - No telemetry, no analytics, no third-party endpoint. File write: - Only under %LOCALAPPDATA%\hermes\ (isolated venv + data directories). - Never touches config.yaml, memory, or any existing skill folder. Environment: - HERMES_HOME, OPENROUTER_API_KEY, LAYA_RUNTIME_SCRIPT, LAYA_CHECKPOINT_DIR, HERMES_PY311. - All read-only; the API key is never logged or echoed.