Hermes Agent Burn Tool

Rust/Burn local inference + training for agent-side temporal signal models.

This is a public, sanitized implementation of a Hermes/OpenClaw-adjacent Burn tool: a GPU-capable Rust binary for running local sequence models beside an agent runtime, plus a training binary for SQLite-backed feature datasets.

The short version: Hermes should not need to call a giant remote LLM for every local signal. This repo shows the lower-level path — small, fast, inspectable models running locally through Burn/WGPU and feeding structured outputs back into an agent loop.

What ships

• joi-burn — WGPU/Metal/Vulkan inference binary for temporal multi-brain/swarm-style signal models.
• joi-train — Burn autodiff training binary for SQLite feature datasets.
• SafeTensors loading helpers.
• Example input and metadata files.
• Public-safe docs and sample commands.
• No private keys, API tokens, trading credentials, local databases, logs, or environment-specific config.

Architecture

Hermes/OpenClaw task
      │
      ▼
feature sequence JSON / local feature DB
      │
      ├── joi-burn  ──► SafeTensors weights ──► structured signal JSON
      │
      └── joi-train ──► SQLite features ───────► new Burn model artifacts

The intended integration pattern is simple: keep the agent orchestration in Hermes, keep heavy/local numeric inference in Rust/Burn, and pass only structured JSON across the boundary.

Why this is interesting for agent infrastructure

• Local-first inference: runs on WGPU-compatible devices instead of requiring a hosted model endpoint.
• Agent-friendly IO: JSON in, JSON out, easy to wrap as a Hermes tool or scheduled worker.
• Burn ecosystem: demonstrates Rust-native model loading/training paths that can sit next to production infrastructure.
• Security boundary: local paths, private datasets, wallets, and credentials stay outside the public repo.
• Composable: can be used as a scoring primitive for trading signals, routing, health models, or any time-series agent heuristic.

Status

Research/experimental. Outputs are model signals, not financial advice and not autonomous trading instructions.

Requirements

• Rust stable
• WGPU-compatible backend
• macOS Metal or Linux Vulkan recommended
• Optional: Hermes Agent or OpenClaw if you want to wire it into an agent runtime

Build

git clone https://github.com/ivanontech/hermes-agent-burn.git
cd hermes-agent-burn
cargo build --release

Release binaries:

./target/release/joi-burn
./target/release/joi-train

Inference quickstart

./target/release/joi-burn \
  --model-dir ./weights \
  --input ./examples/input_sequence.json \
  --output ./burn_brain_signal.json

Expected integration contract:

input_sequence.json  →  joi-burn  →  burn_brain_signal.json

Keep output JSON local if it contains live strategy state.

Training quickstart

./target/release/joi-train \
  --db ./data/features.sqlite \
  --epochs 40 \
  --batch-size 16 \
  --output ./burn_models_v3

Training data/database files are intentionally excluded from Git. Use your own feature database with the schema expected by src/train.rs.

Hermes setup

Clone and build the repo inside your workspace:

git clone https://github.com/ivanontech/hermes-agent-burn.git
cd hermes-agent-burn
cargo build --release

Example local-only environment values:

HERMES_BURN_MODEL_DIR=/path/to/hermes-agent-burn/weights
HERMES_BURN_INPUT=/path/to/input_sequence.json
HERMES_BURN_OUTPUT=/path/to/burn_brain_signal.json

Keep those values in your own Hermes config or .env. Do not commit private paths, live signal outputs, databases, wallets, or secrets.

Files

src/main.rs                  inference binary
src/train.rs                 training binary
examples/input_sequence.json sample feature sequence
weights/README.md            weight placement notes
Cargo.toml                   Burn/WGPU dependencies

Verification

cargo fmt --check
cargo check

For release verification:

cargo build --release
./target/release/joi-burn --help
./target/release/joi-train --help

Security / privacy

Before publishing, this repo was sanitized to exclude:

• .env files;
• API keys/tokens/private keys;
• wallet material;
• local hostnames/IPs and personal absolute paths;
• SQLite databases and JSONL logs;
• private model checkpoints/weights;
• build artifacts.

Run your own scan before redistributing. Report pattern names and filenames only — never paste matched secret values into chat/logs.

grep -RInE 'api[_-]?key|secret|token|private|password|Bearer|sk-[A-Za-z0-9]|/Users/|10\.0\.|\.openclaw|\.hermes' . \
  --exclude-dir target \
  --exclude Cargo.lock

Related

• ivanontech/skunkworks — AI ops dashboard and Alpha Radar cockpit.
• ivanontech/hermes-burn-tool-router — tiny Burn neural router for Hermes tool/control classification.

License

MIT — see LICENSE.