Competitor intel agent — Hermes + Composio + Olostep

A scheduled agent that checks competitor pricing, changelog and blog pages every week and posts what changed, with analysis, to Slack.

• Hermes Agent runs the job on a schedule and does the reasoning.
• Composio gives Hermes its tools through one MCP connection.
• Olostep (via Composio) scrapes each page into structured JSON and researches the "why" behind a change.
• Slack (via Composio) receives one message per run.

Hermes cron (Mon 09:00)
  └─ skill: competitor-intel
       ├─ OLOSTEP_SCRAPE_PAGE × 9      structured JSON per page
       ├─ compare with state/<id>.json  last run's copy
       ├─ OLOSTEP_ANSWER_QUESTION × ≤3  sourced context for the biggest changes
       ├─ save new state
       └─ SLACK_SEND_MESSAGE            one update

Sample output

A real run from 2026-10-13 against Linear, Notion and Asana — see assets/sample-slack-update.txt and the agent's own run report in assets/sample-run-report.md.

Setup

Needs macOS (Apple Silicon) or Linux, a Composio account, an Olostep API key, a Slack workspace, and a model key for Hermes (OpenRouter works).

1. Install Hermes and give it a model

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
echo 'OPENROUTER_API_KEY=sk-or-...' >> ~/.hermes/.env
hermes config set model.provider openrouter

2. Connect Composio to Hermes

Copy your Connect MCP key from dashboard.composio.dev, then:

echo 'COMPOSIO_API_KEY=ck_...' >> ~/.hermes/.env
cat config.example.yaml >> ~/.hermes/config.yaml
hermes mcp test composio        # expect: Connected, Tools discovered

3. Connect Olostep and Slack inside Composio

curl -fsSL https://composio.dev/install | sh
composio login
composio link olostep            # paste your Olostep API key
composio link slack              # approve in the browser

4. Install the skill and watch list

Edit watchlist.json (your competitors, your Slack channel), then:

./install.sh

5. Schedule it

hermes cron create "0 9 * * 1" \
  "Run the weekly competitor intel check by following the competitor-intel skill." \
  --name competitor-intel --skill competitor-intel --deliver local \
  --model anthropic/claude-sonnet-5 --provider openrouter

hermes gateway install           # background service that fires scheduled jobs

Run it once now instead of waiting for Monday:

hermes cron run <job_id>

The first run has nothing to compare against, so it only saves a baseline. Changes show up from the second run.

Files

File	Purpose
watchlist.json	Pages to watch and the Slack channel
schemas.json	What to extract from each page type (pricing, changelog, blog)
skills/competitor-intel/SKILL.md	The agent's instructions
config.example.yaml	Composio MCP block for Hermes
install.sh	Copies the above into ~/.hermes/
seed-baseline.py	Optional: seeds "last week" from Wayback Machine copies

State lives in ~/.hermes/competitor-intel/state/, one JSON file per page.

How the sample output was produced

The sample is one real run, not a week of waiting. seed-baseline.py 20261006 saved the closest Wayback Machine copy of each page (captured 2026-09-22 to 2026-10-06) as the "last run", using the same Olostep extraction. The agent then ran live on 2026-10-13 and reported the difference.

What it costs

• Olostep: structured extraction is 10 credits per page, so about 90 credits per run for 9 pages, plus retries and up to 3 Answers calls.
• Model: Hermes estimated $0.53–$0.69 per run on Claude Sonnet 5 via OpenRouter.
• A run takes about 4–5 minutes.

Known limits

• Extraction varies a little between runs. The JSON comes from an LLM reading the page, so which 8 changelog entries or 6 plan limits it picks can differ. The skill compares by title and plan name and ignores wording to keep false alarms down, but treat a surprising change as a prompt to look, not as fact.
• Geo-routed pages. Some sites localize by visitor location. The skill requests country: "US", retries when a price comes back in another currency, and refuses to report a price change across currencies.
• Script-rendered pages need the wait_before_scraping delay; without it some changelog pages come back empty. Structured extraction can still come back blank on awkward layouts (asana.com/whats-new in the sample watch list did so in about 2 scrapes out of 5), so the skill retries once and then falls back to scraping the page as plain text and reading the entries from that.
• Scraped content is untrusted. The skill tells the agent to treat it as data and to post only to the configured channel. For extra safety, limit the cron toolset to file access plus the Composio MCP server (hermes tools).

Going further

• OLOSTEP_CREATE_MONITOR lets Olostep watch the pages itself and record change events, so the agent only wakes up when something moved.
• OLOSTEP_START_BATCH scrapes thousands of URLs in one call if your watch list grows.