hermes

Run the demos inside Hermes Agent

The same computer-use-agents-demos, with Hermes Agent (Nous Research's open-source agent harness) as the host instead of Claude Code.

Nothing about the SDK or the agents changes. Hermes consumes the same three interfaces these demos already expose: MCP servers, agentskills.io skills, and plain shell commands. This is purely host wiring. Where Claude Code reads .mcp.json, Hermes reads ~/.hermes/config.yaml.

In Claude Code	In Hermes Agent
.mcp.json → hai-agents-demos-qa, ...-extract-anything	~/.hermes/config.yaml mcp_servers: (config.example.yaml)
skills/hai-qa-via-cli, skills/hai-agents	~/.hermes/skills/... (agentskills.io format, same SKILL.md)
qa-cli / extract-cli / counterfeit-cli	identical shell commands

Prerequisites

This repo, installed the usual way (see the root README):

uv sync
cp .env.example .env          # add HAI_API_KEY from https://platform.hcompany.ai/settings/api-keys

1. Install Hermes Agent + MCP support

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash   # Linux / macOS / WSL2
cd ~/.hermes/hermes-agent && uv pip install -e ".[mcp]"              # enable MCP transports

MCP support ships as an extra. Without [mcp] installed, the mcp_servers: block is silently ignored.

2. Register the MCP servers

Hermes' baseline method (per the MCP docs) is editing the mcp_servers: block in ~/.hermes/config.yaml directly. For these demos that is also the safest path, because it sets the required timeout in the same block. A review_web_ui call runs a full cloud browser session that exceeds Hermes' 120 s default, and without the bump it gets killed silently (see Troubleshooting).

Copy the block from config.example.yaml into ~/.hermes/config.yaml, then replace /ABSOLUTE/PATH/TO/agent-sdk-demo with your clone's path (pwd from the repo root prints it). It uses uv run --directory <repo> plus an absolute --env-file, which fixes the one real gotcha: Claude Code's relative --env-file .env won't resolve from Hermes' ~/.hermes working directory.

The hermes mcp add CLI is a convenient alternative (--args captures the rest of the line), but it writes only command/args, so you set the timeout separately:

REPO="$(pwd)"   # run from the repo root
hermes mcp add hai-agents-demos-qa               --command uv --args run --directory "$REPO" --env-file "$REPO/.env" hai-agents-demos-qa
hermes mcp add hai-agents-demos-extract-anything --command uv --args run --directory "$REPO" --env-file "$REPO/.env" hai-agents-demos-extract-anything
hermes config set mcp_servers.hai-agents-demos-qa.timeout 420
hermes config set mcp_servers.hai-agents-demos-extract-anything.timeout 420

Also register the generic platform server

The two servers above are this repo's typed recipes. To drive any H agent, also register the hosted agent-platform server, io.github.hcompai/hai-agents: a streamable-HTTP proxy over /api/v2, hk--key auth, exposing run_agent, wait_for_session, list_agents, send_message, cancel_session, share_session.

hermes mcp add hai-agents-platform --url https://agp.eu.hcompany.ai/mcp --auth header
# set the header to: Authorization: Bearer hk-...   (US endpoint: https://agp.hcompany.ai/mcp)

Config-block form (it stores the key in the file) is in config.example.yaml. run_agent caps its wait at ~24 s, then returns a handle to poll with wait_for_session, so the timeout bump applies for long tasks.

3. Verify

hermes mcp test hai-agents-demos-qa     # MCP handshake + tools/list; no cloud session, no cost

It should report review_web_ui and visual_check. Inside a hermes chat session you can also /reload-mcp and ask "what tools do you have?". You should additionally see extract, describe_picture_of_the_day, and the get_* tools.

Name normalization. Hermes rewrites hyphens and dots to underscores in server names, so hai-agents-demos-qa is surfaced as hai_agents_demos_qa. Tool names like review_web_ui are already valid identifiers and stay as-is.

4. First slice: QA a page

Use review_web_ui to check https://news.ycombinator.com and verify the top story
link works and the page has reasonable accessibility.

Hermes calls the tool, the hai-agents SDK drives a cloud browser, and you get back a structured {verdict, summary, findings}. For a subject with known issues, point it at the intentionally broken page in examples/broken_ui/.

Skills (optional)

Both skills have Hermes-adapted entry docs under hermes/skills/, in agentskills.io format:

mkdir -p ~/.hermes/skills/hai
cp -r  hermes/skills/hai-qa-via-cli ~/.hermes/skills/hai/hai-qa-via-cli   # self-contained
cp -RL hermes/skills/hai-agents   ~/.hermes/skills/hai/hai-agents     # -L dereferences its symlinks (see below)
# then restart Hermes; it re-scans ~/.hermes/skills/

hai-qa-via-cli is forked to invoke the CLI with uv run --directory <repo>. For hai-agents, only the SKILL.md is Hermes-adapted (frontmatter, where h_login.py runs, and writing the key into the demo repo's .env); its references/ and scripts/ are symlinks into the canonical skills/hai-agents/, so there's no duplication and cp -RL copies the real files. (The symlinks need a symlink-aware checkout: macOS, Linux, or WSL2.)

CLI demos

The CLI entry points (qa-cli, extract-cli, counterfeit-cli) are plain shell commands, so Hermes can run them through its built-in shell tools with no registration at all. This is the only way to reach counterfeit_detection, which is CLI-only (local Playwright plus Holo visual compare):

HAI_DEMOS="$(pwd)"   # from the repo root
uv run --directory "$HAI_DEMOS" --env-file "$HAI_DEMOS/.env" counterfeit-cli sweep --genuine-url "https://..."

Notes on the host model

Hermes is model-agnostic (any OpenAI-compatible endpoint; Hermes 3/4 are purpose-built for tool use) and wants at least 64k context for multi-step tool loops. The bar on the host model is low here: the browser-agent reasoning runs in the H cloud via the SDK, and Hermes only decides when to call a tool and reads the result.

Troubleshooting

• The answer says "the tool timed out, so I verified directly" (or isn't a structured ReviewResult). The call exceeded Hermes' default 120 s per-call timeout (mcp_tool.py), so Hermes killed it and fell back to its own web tools. A review_web_ui session runs up to 360 s; set timeout: 420 (step 2).
• The tools don't show up after /reload-mcp. Two common causes: the [mcp] extra isn't installed (step 1), so mcp_servers: is ignored; or you're in an ACP session, where MCP tools may not register because enabled_toolsets is hardcoded (hermes-agent#14986).
• HAI_API_KEY is not set. The server can't read the key. Check that the --env-file path in your config is absolute and points at the repo's .env.

References

• Hermes Agent docs · MCP feature · MCP config reference · Skills system