HERMES
AGENT / v0.9.2

A CLI agent that claims to "autonomously run full software engineering tasks."
Actually good? Or just another hype cycle? — I ran it for 72 hours straight. Here's the honest answer.

rust-core mcp-native 72h-benchmark honest-review

section · 01/04

cat chapter_01.md

// Setup

From brew install hermes to first prompt — 4 minutes 22 seconds total.

benchmark · cold-start

Out-of-box data

cold start → first-successful-task, averaged over 3 runs

install time

42s

Single binary, no Docker, no Python env.

first token

1.8s

Plugged into claude-opus-4-6, no warmup.

first PR merged

4m22s

Ran a fix-a-typo level low-difficulty task.

// verdict +

Cold start is genuinely fast

Compared to OpenClaw's docker + pip flow, it's an order of magnitude faster.

// verdict -

MCP server config is unfriendly

Env vars must be manually stuffed into ~/.hermes/env, docs barely mention it.

trace · hermes run

hermes run "refactor auth module to use pkce"

↓ Real trace (excerpt)

# hermes v0.9.2 · session 42a1
[plan]    "Analyze src/auth/*.ts → find oauth flow → extract to pkce"
[read]    src/auth/oauth.ts       // 214 lines
[read]    src/auth/token.ts       // 88 lines
[think]   "Found implicit flow, migrating to code+pkce, needs new state param"
[edit]    src/auth/oauth.ts       +43 -17
[edit]    src/auth/token.ts       +12 -4
[test]    pnpm vitest auth        PASS 18/18
[commit]  "feat(auth): migrate to oauth2 code+pkce"
[push]    origin feat/pkce-auth   ok

# Total: 3m 14s · 14k tokens · $0.21

benchmark · hermes vs openclaw

72-hour comparison

Same 48 GitHub issues, both agents run independently

tldr

echo $VERDICT

7.8/ 10

Worth installing, not yet worth fully relying on.

+ strong points

• Rust core cold-start is fast
• Trace readability is excellent
• Diff review-friendly, commit messages are decent

- weak points

• Plan phase occasionally skips steps
• Repos over 50k LoC will OOM
• MCP config requires manual env stuffing

install

Want to run it yourself?

Three commands, under 5 minutes to see it do its first thing.

# 1. install
$ brew install hermes-agent/tap/hermes

# 2. auth (have your anthropic api key ready)
$ hermes auth login

# 3. first task
$ cd ~/your-repo && hermes run "add a CHANGELOG.md from git log"

brew-ready opus-4.6 needs-api-key

EOF

exit 0

// thanks

Full trace, PR list for all 48 tasks, benchmark scripts at github.com/lewis/hermes-review