hermes-agent-kit 0.3.5

Production hardening pack for Hermes Agent — per-topic model routing, fallback chains, rate limiting, and cost tracking via gateway hooks

hermes-kit

Production hardening pack for Hermes Agent.

Self-hosted Hermes gateways are powerful but built for single-user setups. Multi-user deployments hit walls: no per-topic model routing, API failures surface as hard errors, one heavy user can burn your API budget with no alert.

hermes-kit fills these gaps with production-grade hooks.

⚠️ How it works: hermes-kit monkey-patches Hermes Agent's internal model resolver at runtime. This is intentionally fragile — Hermes Agent updates may break your setup. We're working on an upstream PR to replace the patch with native hook return values. Until then, test after every Hermes upgrade.

Prerequisites

• Python ≥ 3.11
• Hermes Agent v0.16.0 (pinned)
• Hermes configured with a working model + provider — hermes-kit is a plugin, not standalone. Hermes must be able to respond to messages BEFORE installing hooks.
• A configured gateway (Telegram, Discord, etc.)

Verify Hermes works first

Critical: hermes-kit won't fix a broken Hermes setup. Make sure your gateway responds to messages with the model you want BEFORE installing hooks.

# Set your model (pick one from opencode-go)
hermes /model opencode-go/mimo-v2.5-pro

# Or via config.yaml at ~/.hermes/config.yaml:
cat > ~/.hermes/config.yaml << 'EOF'
model:
  default: opencode-go/mimo-v2.5-pro
  provider: opencode-go

providers:
  opencode-go:
    api_key: OPENCODE_GO_API_KEY
    base_url: https://opencode.ai/zen/go/v1
EOF

# Ensure API keys in ~/.hermes/.env
echo "OPENCODE_GO_API_KEY=sk-..." >> ~/.hermes/.env
echo "GATEWAY_ALLOW_ALL_USERS=true" >> ~/.hermes/.env

# Start gateway
hermes gateway run

# Send a test message on Telegram → verify you get a response
# Only then install hermes-kit

Gateway Management

# Start (with hooks)
hermes-kit gateway run --accept-hooks

# Stop
# Press Ctrl+C in the terminal running the gateway

# Restart after config changes
# Ctrl+C then:
hermes-kit gateway run --accept-hooks

# Check status
hermes-kit status

Install

pip install hermes-agent-kit

🔵 Naming — same project, two names:

Context	Name
PyPI / pip install	hermes-agent-kit
GitHub repo	srmdn/hermes-agent-kit
CLI command	hermes-kit

Why the split? PyPI name matches the repo (hermes-agent-kit). The short CLI alias (hermes-kit) keeps commands terse — hermes-kit install router instead of hermes-agent-kit install router. Same project, same code, two names.

Quickstart

# Install all hooks in one command
hermes-kit install router fallback rate-limiter cost-tracker model-switch

# Verify
hermes-kit doctor

# Start gateway with bridge auto-patched
hermes-kit gateway run --accept-hooks

# If new users get "I don't recognize you":
GATEWAY_ALLOW_ALL_USERS=true hermes-kit gateway run --accept-hooks

Hooks land in ~/.hermes/hooks/<name>/. Hermes discovers them on restart.

Modules

router — Per-Topic Model Routing

Route Telegram topics to different AI models. Finance chat uses Qwen, coding chat uses DeepSeek, everything else falls back to GPT-4o-mini.

Via CLI:

hermes-kit router set-default --model opencode-go/qwen3.6-plus
hermes-kit router add 42 --model opencode-go/deepseek-v4-pro
hermes-kit router show

Via YAML (~/.hermes/hooks/router/topic_router.yaml):

default:
  model: "opencode-go/qwen3.6-plus"

topics:
  "42":
    model: "opencode-go/deepseek-v4-pro"

Multi-provider — route specific topics to native providers:

hermes-kit router add 42 --model gpt-4o --provider openai
hermes-kit router add 7 --model claude-sonnet-4-20250514 --provider anthropic

Hermes resolves API keys from ~/.hermes/.env (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.). See providers guide for all supported providers and model IDs.

fallback — Automatic Fallback Chain

Define a chain of models to try when the primary fails.

Via YAML (~/.hermes/hooks/fallback/fallback_chain.yaml):

chains:
  global:
    - "opencode-go/deepseek-v4-pro"     # primary
    - "opencode-go/kimi-k2.6"      # fallback
    - "opencode-go/qwen3.6-plus"          # last resort

After a failure, call hermes_kit.bridge.retry_with_fallback(session_key) to advance to the next model.

rate-limiter — Per-User Rate Limiting

Prevent a single user or chat from draining your API budget.

Via YAML (~/.hermes/hooks/rate-limiter/rate_limits.yaml):

limits:
  global:
    max_messages_per_window: 100
    window_seconds: 3600
  per_user:
    "123456789":
      max_messages_per_window: 50

Rate limiter enforces limits at the bridge level — exceeding users get RateLimitExceeded on every request until their window resets. Set window_seconds to control reset timing.

cost-tracker — Real-Time Cost Tracking

Track token costs per session and alert when thresholds are exceeded.

Via YAML (~/.hermes/hooks/cost-tracker/cost_tracker.yaml):

alert_threshold_usd: 1.0

Set to 0 to disable alerts but continue tracking.

Docs

• Quickstart — agent-driven and manual install
• Providers — supported AI providers and model lists
• Manual setup per module:
  • Router — per-topic model routing + /route command
  • Fallback — automatic retry chains
  • Rate Limiter — per-user quotas
  • Cost Tracker — budget alerts
• Troubleshooting — common issues

License

MIT