✨ A cheerful add-on kit for Hermes Agent · MIT licensed

Give your agent a memory, a map, and a health plan.

Long agent sessions bury their best clues in scrollback. Harness Plus keeps the receipts with Task Canvas evidence trails, local semantic recall, share-ready skills, and an optional provider-backed minion bridge. Every piece is operator-controlled. 🧸

7+pluggable pieces 🧩
Localstate by default 🏠
0changes to Hermes core 🪶
MITdo-what-you-love license 📜
🪦 The scrollback graveyard

Your agent works hard.
Its best clues deserve better than line 4,000.

A tool call surfaces the golden hint. A test log answers the big question. Ten minutes later, it's all buried — and everyone pays for it.

😵‍💫

Forgotten clues

The crucial output scrolls away, and the session quietly loses the shape of the work.

🔁

Repeated dead ends

Without a map of what failed, the agent happily tries the same broken path again.

🤷

Unverifiable answers

"Trust me" is not a citation. Final answers should point back to real evidence.

🎁 What's in the box

Seven little helpers, one happy harness.

Every piece is optional and independent. Take one, take all — Hermes Agent itself is never modified.

🗺️

Context Canvas

A tiny library that keeps short JSON nodes plus evidence refs — so every finished note has a path back to the facts.

Read the origin story →
🛰️

Canvas MCP sidecar

Native MCP tools for Hermes — start a map, recover recent canvas IDs, or attach evidence and its node in one atomic call.

🔎

Qdrant recall kit

Index selected skills and recent sessions into your own local Qdrant — with dry-run previews and secret-pattern redaction.

🩺

Recall watchdog

Scheduled health checks that stay quiet on green and get specific on failure — plus bounded Docker restart helpers.

📚

Skills & checklists

Drop-in Hermes skills and ops checklists that teach the habits: keep the map, find it again, share cleanly.

🧭

Delegation calibrator

Optional complexity × Bayesian evidence for direct versus luna_max — with Baton and the main agent still in charge of judgment.

Read the routing guide →
🤖

Prime minion bridge

Run Prime Agent's coding loop with Hermes-owned Codex OAuth, explicit route readback, and optional resumable transcripts.

Meet the minion bridge →
⚡ Quick start

From zero to first canvas in about a minute.

Three copy-paste steps. The full install guide covers MCP wiring, retired-plugin cleanup, and Qdrant helpers.

Grab the toolbox

Clone the repo and hop in.

# clone & enter git clone https://github.com/phenomenoner/\ hermes-agent-harness-plus.git cd hermes-agent-harness-plus

Prove it works

Install and run the smoke tests.

# install + smoke tests python -m pip install -e '.[mcp]' pytest python -m pytest -q

Start a canvas 🎉

Give your first task a memory.

# first Task Canvas PYTHONPATH=packages/context-canvas \ python -m context_canvas.cli start \ --session-id demo \ --goal "Keep evidence for a long Hermes task"
🧭 Our promises

Cute on the outside, careful on the inside.

🏠 Local state by default 🪶 Zero changes to Hermes core 🛟 Failure posture matches the boundary 🔍 Dry-run previews before any indexing 🧾 Every claim links to evidence
💬 Tiny FAQ

Good questions, honest answers.

Does it modify Hermes Agent?

Nope! Everything is an optional outer layer — packages, sidecars, skills, helpers, and checklists. Remove any piece and Hermes keeps running exactly as before.

Where does my data go?

Nowhere. Canvases live under ~/.hermes/context-canvas, and Qdrant recall binds to 127.0.0.1 by default. Session indexing skips system prompts and tool outputs, redacts common secret patterns, and always supports --dry-run previews.

Do I need all six pieces?

Not at all — they're independent. Lots of value starts with just the Task Canvas CLI. Add the MCP sidecar, Qdrant recall, or the optional delegation calibrator whenever your workflow asks for them.

What do I need to run it?

Python 3.10+, plus a working Hermes Agent install for the MCP sidecar. The Qdrant helpers want a local Qdrant (Docker works great) and fastembed. Details live in the install guide.