agentos.guide › blog

Hermes Fable 5: Put Claude Fable 5.1 Behind The Hermes Agent In Your Agent OS, With 5 Copy-Paste Prompts

By Julian Goldie · 4 November 2026 · agentos.guide

Hermes Fable 5 is the setup that puts Anthropic's Claude Fable 5.1 behind the Hermes agent in your Agent OS, and your agent can do most of the wiring for you.

You give it a prompt, it runs the commands, and it shows you the result.

Below is the quick start, five copy-paste prompts and a daily workflow that keeps Fable on the hard thinking.

If you'd like my team to plan your agent stack with you, book a free SEO strategy session.

The Quick Start

This uses the official experimental plugin, claude-subscription-directsdk, from Nous Research.

It runs the Claude Code CLI as a model provider on your Claude plan, so there's no API key to paste.

npm install -g @anthropic-ai/claude-code
claude auth login
hermes plugins install claude-subscription-directsdk
hermes model

In the picker, choose Claude Subscription DirectSDK (Experimental), then Fable.

You'll need Hermes Agent 0.21.4 or newer and Python 3.10 or newer, per the README.

Fable 5.1 is pinned as claude-fable-5-1[1m], with a 1,000,000-token context window.

For the model itself, what launched and what it's good at, read my Claude Fable 5 guide.

This page is only about the Hermes wiring.

Prompt 1: Get Your Agent To Install Everything

Paste this into your Agent OS chat.

Set Hermes up to run on Claude Fable through the official plugin.
1. Check that the claude CLI is installed. If not, run:
   npm install -g @anthropic-ai/claude-code
2. Run claude update so the CLI supports the newest models.
3. Run claude auth status. If I'm logged out, start claude auth login
   and give me the link to approve.
4. Run: hermes plugins install claude-subscription-directsdk
5. Show me hermes --version and confirm it is 0.21.4 or newer.
Report each step as done or failed.

The one step you do yourself is approving the Claude login in your browser.

Everything else the agent handles.

Prompt 2: Make Fable The Main Brain

Update ~/.hermes/config.yaml so the main model block reads:
model:
  provider: claude-subscription-directsdk-experimental
  default: fable
Don't change any other section.
Then show me the model block and start a new session to confirm
the welcome banner names Fable.

The fable short name resolves to the pinned Fable 5.1 route.

Hermes keeps its own loop, tools, approvals and compaction on top.

Get the Agent OS with Fable already in the stack

Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members running Hermes.

→ Join the AI Profit Boardroom

Prompt 3: Check Your Plan Before Long Sessions

Your plan decides what Fable costs you.

Commit fdbf4225d2 in the plugin repo applies Anthropic's rule: Fable is included on Max and Ultra, and runs on usage credits on Pro and on standard Team and Enterprise seats.

Run /usage and tell me my current plan limits.
Then open the model picker and tell me whether the Fable row carries
a "usage credits" note.
If it does, warn me before any long Fable session.

Then do one thing by hand: switch off extra usage in your Claude account if you don't want overage.

That's straight from the plugin README.

If /usage says it can't fetch limits, check that claude_code_telemetry is still on, because switching it off skips that fetch.

Prompt 4: Send Subagents To A Lighter Model

Subagents copy the parent model unless you tell them otherwise, says the delegation guide.

That means every delegate_task child would run on Fable too.

Add this to ~/.hermes/config.yaml without touching the model block:
delegation:
  provider: claude-subscription-directsdk-experimental
  model: sonnet
Then run one small delegated task, such as summarising three files,
and tell me which model the subagent used.

Fable plans and checks.

Sonnet does the batch work.

Prompt 5: Move Every Cron Job Off Fable

Cron picks a model in this order: the job's own pin, then cron.model, then your main model.

The cron documentation spells that out.

So once Fable is your main model, every unpinned job follows it.

Run hermes cron list and show me every job with its model.
Set a fleet default for unpinned jobs:
  hermes config set cron.model_provider claude-subscription-directsdk-experimental
  hermes config set cron.model haiku
For any job that writes a long report, pin it to sonnet instead:
  hermes cron edit <job_id> --provider claude-subscription-directsdk-experimental --model sonnet
Show me the list again when you're done.

Hermes's own code agrees with this split: models.py refuses to pick Opus or Fable as a silent default, because it would carry into every cron job.

Check The Picker Lists Your Real Models

Since commit 4220bbaadb, the picker asks Claude Code for your account's own model list.

That means it shows what your plan can actually run, and older pinned models stay selectable underneath.

Open the Hermes model picker for the Claude Subscription DirectSDK provider.
List every model it shows, with any "usage credits" or "unpinned" note.
Tell me which row the fable short name maps to.

If the list looks short, the likely cause is claude_code_telemetry switched off.

The README says that setting fetches the feature flags that unlock models, so leave it on unless you have a reason not to.

When To Open A Fable Session, And When Not To

Open a Fable session when the job is long, has many steps, or needs several documents read together.

Think quarterly planning, a full proposal, a policy rewrite, or a research brief that ends in a finished document.

Don't open one for a two-line reply, a quick lookup or a daily summary.

Those belong to the lighter models you set up in Prompts 4 and 5.

Each /model switch also resets the prompt cache, per the Hermes FAQ, so avoid flipping back and forth mid-session.

Start a fresh session on Fable instead, or hand the small task to a subagent.

The Backup Route: Hermes's Built-In Anthropic Provider

Not on Max or Ultra, or need a path that isn't experimental?

Use the native Anthropic provider instead.

export ANTHROPIC_API_KEY=your-key
hermes chat --provider anthropic --model claude-fable-5-1

The providers guide says an API key is billed per token to the key's organisation, apart from any Claude plan.

The OAuth login on this route is narrower: it only works on Claude Max with purchased extra usage credits, it never draws on the base Max allowance, and Claude Pro can't use it.

Ask your agent to set this up only if Prompt 1 failed or your plan doesn't suit the plugin.

A Daily Workflow For An Estate Agency

Here's how a small lettings and sales office could run it.

1. At 7am, a cron job on Haiku reads new enquiry emails and sorts them by viewing, valuation or complaint.
2. It drops the sorted list into the office chat.
3. At 9am, the manager opens the main chat on Fable with the week's listings, notes and the sorted enquiries.
4. Fable plans the day's viewings and drafts replies for the tricky enquiries.
5. It hands each listing write-up to a Sonnet subagent with the house style.
6. The manager checks the drafts and sends them.
7. On Friday afternoon, the manager asks Fable for a weekly review of every enquiry and viewing, with follow-ups for Monday.

Fable does the planning and the hard replies.

The lighter models do the sorting and the write-ups.

What Fable 5.1 Brings To The Agent

Anthropic released claude-fable-5-1 on 1 September 2026, and its model page lists a 1M-token window, 128K output and thinking that is always on.

Fable 5 is now a legacy model.

Anthropic's models overview says to start with Opus 5.5 for most work and use Fable 5.1 for demanding reasoning and long-horizon agentic jobs.

That's why Fable belongs in the main chat and not in every helper.

Default effort on Fable 5.1 is high, and its reliable knowledge runs to June 2026.

Anthropic's what's new page flags a few breaking changes from Fable 5.

Forced tool use now returns an error, and only Fable 5.1 or Mythos 5.1 can read Fable 5.1's thinking blocks.

Neither matters much for day-to-day Hermes use, because Hermes owns the tool loop.

It does matter if you've pinned claude-fable-5 somewhere in an old config, so search your profiles and move those lines to the fable short name.

Fable 5 stays available for now, with retirement no sooner than 9 June 2027.

One more practical note: the plugin never swaps you onto a different model when Fable is refused.

It shows the error as it comes back from Claude Code, so you always know which model answered.

Three Mistakes To Avoid

• An API key left in your shell. The plugin refuses inherited API-key and custom-endpoint overrides and names the clashing variable, so unset it first.
• Expecting a silent fallback. If your plan needs usage credits, the plugin shows "Fable 5.1 requires usage credits" and does not swap models for you.
• Treating it as production-ready. The README calls it a review build, so keep a person on the final step.

For the rest of your model line-up, see the best Hermes Agent model guide.

Want help planning which jobs get Fable?

Book a free SEO strategy session and my team will map your agent jobs to the right models with you.

→ Book your free call

Ready to run the full system? Join the AI Profit Boardroom and install the Agent OS.

FAQ

Can my agent install the Fable plugin itself?

Yes, apart from the Claude Code login, which opens a browser for you to approve. The agent can run every other command and report back.

What should the model block say after the switch?

provider: claude-subscription-directsdk-experimental and default: fable. The fable short name resolves to claude-fable-5-1[1m].

Why give subagents a different model from the main chat?

Subagents copy the parent model unless you set delegation.model. Pointing them at a lighter Claude model keeps Fable for the planning.

What if my plan bills Fable to usage credits?

On Pro and standard Team or Enterprise seats it does. Either accept the credits, switch off extra usage in your Claude account, or use the built-in Anthropic provider with an API key.

Is this the same as the Claude Fable 5 launch guide?

No. The launch guide covers the model itself. This page covers wiring Fable into the Hermes agent inside your Agent OS.

Final word on Hermes Fable 5

Paste Prompt 1 into your agent now, and if you want the whole system ready to go, join the AI Profit Boardroom and run Hermes Fable 5 inside the Agent OS today.

Related guides on agentos.guide

→ The Best Hermes Agent Model, And The Five Model Lanes I Route Inside My Agent OS

→ Hermes Delegation: Run Parallel Subagents In Your Agent OS With 5 Copy-Paste Prompts

→ Hermes Agent Cron Job: Schedule, Stop And Recover Jobs In Your Agent OS With 5 Copy-Paste Prompts

→ Hermes Compaction: Five Prompts To Tune The Trigger, The Tail And The New prepare_ahead In Your Agent OS

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

← all guides on agentos.guide