Hermes is NousResearch’s terminal-focused CLI agent. The Lowkey pack installs it and points it at bedrockify, a local OpenAI-compatible proxy that translates Hermes’s API calls into Bedrock InvokeModel requests. You get Hermes’s workflow — including its self-improving skills and learning loop — with Claude or other Bedrock models as the backing inference, and no API keys to manage.
Experimental pack. Hermes is under active development by NousResearch. Expect upstream changes between Lowkey versions.
What makes Hermes different
- Self-improving skills. Hermes has a built-in learning loop that improves the skills it uses over time, letting the agent get better at tasks it repeats.
- Lightweight. Hermes runs on a
t4g.medium with no data volume. It’s the right choice when you want a capable terminal agent without the resource overhead of OpenClaw.
- Bedrock via bedrockify. Hermes expects an OpenAI-compatible API. The pack configures it to talk to bedrockify at
http://localhost:8090/v1, which translates its calls into Bedrock requests using the instance’s IAM role.
- OpenAI-style model IDs. Hermes uses model IDs in the format
anthropic/claude-opus-4.6, not Bedrock-style IDs like us.anthropic.claude-opus-4-6-v1. Bedrockify handles the translation.
- No gateway, no memory service. Unlike OpenClaw, Hermes is a pure CLI binary. Nothing runs in the background between sessions.
Compatible profiles
| Profile | IAM permissions | Instance | Use case |
|---|
builder | AdministratorAccess | t4g.medium | Build apps, deploy infra |
account_assistant | ReadOnlyAccess + Bedrock | t4g.medium | Architecture review, cost analysis |
personal_assistant | Bedrock only | t4g.medium | Writing, research, coding help |
Prerequisites
- AWS CLI configured with admin access in a dedicated sandbox account
- Amazon Bedrock model access enabled in your target region (default:
us-east-1)
- No API keys, Docker, or interactive login needed
Install
curl -sfL install.lowkey.run | bash -s -- -y \
--pack hermes \
--profile builder
--skip-setup to suppress interactive prompts. It writes ~/.hermes/config.yaml and ~/.hermes/.env, pointing Hermes at bedrockify on localhost.
Connect and use
After the stack deploys, SSM into the instance and run hermes directly:
# Open a session
aws ssm start-session --target <instance-id>
# Verify the install
hermes --version
# Start an interactive Hermes session
hermes
Configuration options
| Flag | Default | Description |
|---|
--region | us-east-1 | AWS region for Bedrock |
--hermes-model | anthropic/claude-opus-4.6 | Model ID in OpenAI-style format (via bedrockify) |
--bedrockify-port | 8090 | Port where bedrockify listens |
Use --hermes-model to change the model, not the generic --model flag. The generic --model flag carries Bedrock-style IDs (e.g. us.anthropic.claude-opus-4-6-v1) from the bootstrap dispatcher, which Hermes does not accept. The pack ignores --model entirely and reads --hermes-model instead.
--hermes-model:
| Model | ID |
|---|
| Claude Opus 4.6 (default) | anthropic/claude-opus-4.6 |
| Claude Sonnet 4.6 | anthropic/claude-sonnet-4.6 |
Resource requirements
| All profiles |
|---|
| Instance type | t4g.medium |
| Root volume | 40 GB |
| Data volume | 0 GB |
Notes and limitations
- Hermes is a pure CLI pack — no background gateway service runs between sessions. Each
hermes invocation is a fresh process.
- Hermes has no built-in persistent memory between sessions the way OpenClaw does. If you need cross-session continuity and multi-channel access, use OpenClaw instead.
- The
--hermes-model flag is Hermes-specific. You cannot use Bedrock ARNs or cross-region inference IDs directly — bedrockify maps the OpenAI-style ID to a Bedrock model ID internally.
- Tearing down the stack leaves no lingering state to clean up — the data volume is 0 GB and the instance has no persistent agent storage.
