Hermes Agent Provider: Anthropic API (Claude family — Slack-bot, Computer-Use)

Why Anthropic

Anthropic’s Claude family is the most-used provider on Hermes Agent today. Three reasons:

1. The model family is wide. Haiku 4.5, Sonnet 5, Opus 4.8 — three tiers with predictable price/capability ratios. Operator can pin haiku-class for cheap jobs, sonnet-class for normal work, opus-class for the long thinking chain.

2. Reasoning quality is high for structured output. Hermes Agent’s skills produce structured markdown, JSON, or yaml. Claude follows formatting instructions tightly; other providers drift on long outputs.

3. Slack-bot class works well. Claude is the most-used as a Slack-bot behind the Slack gateway. Slack message rendering, threaded replies, emoji handling — all work cleanly.

Computer-Use is Anthropic’s other flagship use case. Hermes supports it via --toolsets computer-use (a research preview). If you’re running automated UI tests, this is the provider to wire.

The 4-line setup

The Anthropic provider on Hermes:

1. export ANTHROPIC_API_KEY=<key> from https://console.anthropic.com/settings/keys.
2. hermes config set provider.anthropic.api_key "$ANTHROPIC_API_KEY" to persist it.
3. Pin a default model: hermes config set provider.anthropic.default_model "claude-haiku-4-5-latest" (or pick the class that fits your job shape).
4. hermes chat --provider anthropic -q "Reply with OK" — verify it works in 1-3 seconds.

If step 4 times out, the failure is usually step 1 (key not exported), step 3 (model name stale), or the live API is having an outage (check status.anthropic.com). The outage case is the easiest to miss: the daemon swallows the 503 error, returns a generic timeout, and the operator thinks the install is broken. Always check status first when the chat just hangs.

Model pin: -latest vs date-snapshot

claude-haiku-4-5-latest always points at Anthropic’s current best Haiku; claude-haiku-4-5-20250514 is a date-snapshot and never updates. Pin by -latest to ride the upgrade treadmill for free.

# Always use -latest
hermes config set provider.anthropic.default_model "claude-haiku-4-5-latest"

# Avoid date-snapshot pins (they go stale in weeks)
hermes config set provider.anthropic.default_model "claude-haiku-4-5-20250514"   # DON'T

Re-pin monthly because model names get renamed on each major release. The live matrix at https://agenticbotsitter.com/websites/ai-model-matrix/model-matrix.html is the canonical reference.

The Claude family quick-reference

Tier	Current name	Input $/Mtok	Output $/Mtok	Use for
Cheap	claude-haiku-4-5-latest	$1	$5	First-pass drafting, classification, low-stakes output
Balanced	claude-sonnet-5-latest	$3	$15	Default for medium jobs
Frontier	claude-opus-4-8-latest	$5	$25	Long-context reasoning, Computer-Use, when correctness matters more than cost

The full live table is at the matrix URL above. Refresh before any pin change.

Slack-bot and Computer-Use use cases

For a Slack-bot, the cheapest setup is haiku-class pinned; the bot has short messages and doesn’t need a frontier model unless the conversation threads into long reasoning. Pin to haiku and override per-thread via --model claude-sonnet-5-latest when the thread warrants.

For Computer-Use (UI automation via Hermes + Claude), pin to opus-class. Computer-Use reasoning chains are long; the cheaper models run out of reasoning budget by mid-task. Operators running screenshot loops on Claude Computer-Use pin to opus and accept the cost.

Cross-provider positioning

Anthropic is rarely the only provider on a real Hermes install. Three roles it plays in a multi-provider setup:

• Primary chat (most common). Anthropic first for general-purpose drafts — the Claude family is the cheapest reliable model for structured output.
• Fallback (second slot). OpenAI or Gemini as the safety-net. If Anthropic has an outage (status.anthropic.com), the fallback catches the load.
• Computer-Use-only (rare). When you specifically need Computer-Use, pin that single skill to Anthropic Opus and let the rest of the install default to a cheaper provider.

A complete three-provider setup typically looks like:

~/.hermes/config.yaml
  provider_fallback:
    - anthropic_api
    - openai
    - gemini

The first slot doesn’t have to be Anthropic — pick whichever is cheapest per token for the dominant job shape. For ABS today it’s Anthropic first; for a heavy image-gen install, it’s OpenAI first.

What not to do

• Don’t pin a date-snapshot model — -latest only.
• Don’t trust Anthropic’s dashboard usage display — it’s hourly-sampled. Realtime: curl /v1/organizations/usage.
• Don’t put OpenRouter-routed Anthropic first in the fallback chain. Direct Anthropic is faster than aggregator-routed Anthropic.
• Don’t forget the soft-cap (provider.anthropic.monthly_cap_usd 50) on shared installs.

Verification checklist

#	Command	What it proves
1	hermes config show --provider anthropic	Provider configured
2	hermes chat --provider anthropic -q "Reply with OK"	API key valid
3	hermes chat --provider anthropic --model claude-haiku-4-5-latest -q "Reply with OK"	Model pin resolves
4	curl https://status.anthropic.com/api/v2/status.json	Live status: operational
5	monthly_cap_usd set in config	Soft-cap configured

Sources

• Anthropic console
• Hermes Agent provider docs
• ABS live matrix: AI Model Matrix
• ABS companion: Anthropic Provider on ABS: Quick Setup and the Cheap Mistakes

Last verified: 2026-07-14 against Hermes v0.18.2 (2026.7.7.2). Anthropic’s published pricing snapshot 2026-07-14.