Featherless for Hermes Agent: Plans, Compatibility, and Setup

Quick decision

• Chat ($25/month) gives unlimited interactive inference within a 32K context and a 4-unit concurrency budget. That 32K ceiling is below Hermes Agent’s documented 64K minimum, so I do not recommend Chat as a self-hosted Hermes backend. It remains fine for ordinary API chat clients.
• Developer / token-based provides up to 256K context and prepaid per-token credits. This is the appropriate Featherless API plan for a locally installed or VPS-hosted Hermes Agent.
• Managed Hermes Agent is available through Featherless’s Agent Marketplace on eligible plans. Choose this if you want Featherless to host the agent and inference together — no local Hermes install, no separate endpoint config.
• Self-hosted Hermes connects through a custom OpenAI-compatible endpoint at https://api.featherless.ai/v1.

Last verified: August 2, 2026 against Featherless pricing, Featherless Plans, Concurrent Unit Limits, Managed Hermes, Request Pricing, and the Hermes providers docs. Plans, prices, and the model catalog drift — re-check the live pages before paying or debugging.

Already decided? Self-hosted → Configure Hermes. Managed → Featherless Managed Hermes.

Featherless is an OpenAI-compatible inference host for open models — one API key for tens of thousands of models, no per-lab accounts. But there is one compatibility fact you must internalize before you wire it into Hermes Agent: Hermes Agent requires a model with at least 64,000 tokens of context, and rejects smaller windows at startup. Featherless’s entry Chat plan tops out at 32K. So the cheap plan that looks perfect for “chat” is the wrong tier for a self-hosted Hermes backend.

The practical split is between two routes:

• Self-hosted Hermes (you run Hermes locally or on a VPS) → use a Featherless plan that provides ≥64K context. That means the Developer or token-based plan (up to 256K). You point Hermes at https://api.featherless.ai/v1 as a custom endpoint.
• Managed Hermes (Featherless runs the agent for you) → launch Hermes from the Featherless Agent Marketplace. Featherless is already wired as the backend; no local install, no endpoint config.

Featherless positions itself as built by researchers contributing to RWKV, a Linux Foundation project — serverless hosting for open models rather than a closed lab. Featherless home, verified August 2, 2026

Contents

1. The plans at a glance
2. Why Chat’s 32K conflicts with Hermes
3. What “unlimited tokens” and “concurrent units” really mean
4. Subscription or API access?
5. Models, modalities, and exclusions
6. Which tier should you buy?
7. Two routes: managed vs self-hosted
8. Featherless Managed Hermes Agent
9. Commercial caveats before checkout
10. Before you set up Hermes Agent
11. Configure Featherless in Hermes as a Custom Endpoint
12. Command map
13. Smoke-test the connection
14. Troubleshooting
15. Security and cost controls
16. Keep this setup current

The plans at a glance

Featherless’s public pages use slightly different product names and starting prices between the home page and the plans documentation. The table below reflects both as of August 2, 2026; treat the live checkout as authoritative.

The Chat and Developer cards on the home page are the consumer-facing shape most readers will buy. The plans documentation describes the Token-Based Business tier at $25+ with 8 concurrent units and one agent sandbox — that $25+ tier is the per-token “Developer” shape. Featherless home, verified August 2, 2026; Featherless Plans docs, verified August 2, 2026

Why Chat’s 32K conflicts with Hermes

This is the single most important thing in the guide. The current Hermes Agent documentation states:

Hermes Agent requires a model with at least 64,000 tokens of context. Models with smaller windows cannot maintain enough working memory for multi-step tool-calling workflows and will be rejected at startup.

Featherless Chat caps at 32K context. A Hermes Agent pointed at a 32K-capped plan — or configured with context_length: 32000 — is below Hermes’s required minimum and will not run its agent loop. So:

• Do not use Chat as a self-hosted Hermes backend. It is fine for an ordinary OpenAI-compatible chat client, but not for Hermes Agent.
• Use Developer / token-based (256K) for self-hosted Hermes. That clears the 64K bar with room to spare.
• If you only want an agent and do not want to run Hermes yourself, the managed Hermes route (below) bypasses the context question entirely — Featherless runs the agent on eligible plans.

Featherless home, tier cards, verified August 2, 2026; Hermes Agent providers — “Minimum context: 64K tokens,” verified August 2, 2026

What “unlimited tokens” and “concurrent units” really mean

The headline on the Chat tier is “unlimited tokens” — that is the model of the offer, but two limits still apply.

1. A hard concurrent-units budget. Every model carries a per-request concurrent-unit cost. With Chat’s 4 units, you can run four small-model requests at once, two larger-model requests, or one very-large-model request — or any mix that adds up to four. Anything that would push you over returns HTTP 429. Featherless Concurrent Unit Limits, verified August 2, 2026
2. A model-size limit. Models are bucketed by parameter count:

The exact cost is set per model class and is shown on each model’s page; the table above is the published rule of thumb. Featherless Concurrent Unit Limits, verified August 2, 2026

There is no published monthly token ceiling on Chat and no published monthly ceiling on Developer’s credit pool, but Developer is billed per token so its ceiling is whatever the credit balance buys. Developer explicitly allows unused credits to roll over. Featherless home, “Developer” tier card, verified August 2, 2026

The plans are sold as fixed monthly subscriptions, not annual prepay. Featherless’s pricing section says “Billed monthly. Cancel anytime.” Featherless home, pricing card, verified August 2, 2026

Subscription or API access?

Featherless offers two ways to pay, and they are easy to conflate:

• Monthly subscription (this guide’s Chat / Developer). One API key, billed monthly, with the concurrent-unit budget described above. Best when you want predictable cost and one key for the whole open-model catalog.
• Pay-as-you-go / per-request API. The Per-Request business tier bills per successful request by model price and token usage, with 100 concurrent units and no model-size limit. Best for bursty or production fleets where a flat monthly plan would leave capacity on the table.

Featherless is always an API key behind an OpenAI-compatible endpoint; the choice is flat-monthly (Chat/Developer) vs metered (Per-Request). For light, sporadic use, compare the per-request option against the $25 Chat subscription — Chat becomes more attractive once you use it frequently enough to value predictable monthly pricing and unlimited tokens within its concurrency limits.

Models, modalities, and exclusions

Featherless positions itself as one key for tens of thousands of open models. The live catalog counted 43,730 on August 2, 2026, and the managed-Hermes docs reference 30,000+; treat all as approximate. Featherless model catalog, verified August 2, 2026; Featherless Managed Hermes, verified August 2, 2026

Models on the catalog include DeepSeek (V3.2, V4-Flash, V4-Pro), GLM (5.1, 5.2), Kimi (K2.6, K3), Qwen (3.5, 3.6, Qwen3-Coder), Llama 3.x, Mistral, Gemma 4, GPT OSS (20B, 120B), and a long tail of roleplay and chat fine-tunes. Model IDs use the Hugging Face-style org/name slug (for example Qwen/Qwen2.5-7B-Instruct, MiniMaxAI/MiniMax-M3). Featherless model catalog, verified August 2, 2026; Featherless Quickstart guide, verified August 2, 2026

Two exclusions to keep in mind:

• Chat and Developer are consumer plans. Featherless’s Terms of Service say individual plans are “for interactive use or prototyping and experimentation by the purchaser” and that misuse (resale, batch, multi-user production) “will have their subscription terminated and no refund will be provided.” For customer-facing or multi-user workloads, use a Business tier. Featherless Terms of Service, Section 3, verified August 2, 2026
• Context length is a hard ceiling, not a soft hint. Chat caps at 32K — below Hermes’s 64K minimum. Developer goes up to 256K. Treat 256K as the working ceiling for self-hosted Hermes. Featherless home, tier cards, verified August 2, 2026

A few model-page details worth flagging because they affect Hermes setup:

• Vision input. Several models in the catalog advertise Vision (for example MiniMaxAI/MiniMax-M3, google/gemma-4-31B-it). Whether image input is honored on your plan depends on the per-model settings, not the global tier.
• Tool calling. Many of the coding and productivity models in the catalog advertise tool calling (“Tools”). Hermes uses OpenAI-style tool_choice: "auto" by default, so a tool-capable model is what you want if you plan to let Hermes call functions. Featherless model catalog, “Tools” tag, verified August 2, 2026
• No native Hermes model name. Hermes does not ship a built-in featherless provider or a Featherless-specific model catalog. The Custom Endpoint path takes the model id as a string.

Which tier should you buy?

For self-hosted Hermes: Developer / token-based

This is the only Featherless API plan that clears Hermes’s 64K minimum. It gives 256K context, prepaid per-token credits, and 1 included agent sandbox. Pick this if you run Hermes locally or on a VPS and want the open-model catalog as your backend.

Chat: not for Hermes Agent

Chat is the right answer for an ordinary API chat client — personal chat, roleplay, writing, trying many open models — but its 32K ceiling is below Hermes’s required minimum, so do not point a self-hosted Hermes at it.

Managed Hermes: skip the install

If you do not want to run Hermes yourself, the managed route (next section) hosts the agent on eligible Featherless plans. No local install, no endpoint config.

A safe buying rule for self-hosted use: start on the token-based plan, monitor the billing dashboard for a week, and adjust the monthly credit amount once you see real usage. Chat is “Cancel anytime.” Featherless home, pricing card, verified August 2, 2026; Featherless Terms of Service, Section 4, verified August 2, 2026

Two routes: managed vs self-hosted

Route	Best for	Setup
Featherless-managed Hermes	Someone who wants an always-on agent without managing a server	Launch Hermes from the Featherless Agent Marketplace
Self-hosted Hermes + Featherless inference	Someone already running Hermes locally or on a VPS	Configure Featherless as a custom OpenAI-compatible endpoint

Both use Featherless as the model backend. The difference is only who runs the Hermes process.

Featherless Managed Hermes Agent

Featherless offers a fully managed Hermes Agent — a hosted agent environment with Featherless already wired as the model backend. You get persistent memory, skills, tool access, and channel connections (including Telegram) without installing or maintaining Hermes yourself. Agent sandboxes are available on plans above the $25 entry tier. Featherless Managed Hermes, verified August 2, 2026

To launch:

1. Go to featherless.ai/account/agents and click Browse Marketplace.
2. Find Hermes Agent and click Launch.
3. Pick a model from the dropdown (the default is MiniMaxAI/MiniMax-M2.5; token-based plans expose NousResearch/Hermes-3-Llama-3.1-8B, MiniMaxAI/MiniMax-M2.5, and others).
4. Your sandbox spins up in seconds; the gateway dashboard opens at https://[port]-sbx-[your-id].sandbox.featherless.ai.

From the dashboard you can chat, browse sessions, edit config.yaml, manage API keys, connect Telegram/Discord/Slack/WhatsApp/Signal, and view cron jobs. This is the lowest-friction way to get a Hermes Agent on Featherless, and it sidesteps the 64K context question entirely because Featherless runs the agent on eligible plans.

Commercial caveats before checkout

The points that affect a purchase decision:

• The current month is non-refundable. Section 4 says renewals can be canceled at any time, but the current month is not refunded. Cancel before renewal. Featherless Terms of Service, Section 4, verified August 2, 2026
• Plans and models can change. The service is “in beta”; Featherless “may change or discontinue any part of the Services,” including de-listing models. Featherless Terms of Service, Section 1, verified August 2, 2026
• Individual plans are not for resale or multi-user production. Misuse terminates the subscription with no refund. Featherless Terms of Service, Section 3, verified August 2, 2026
• You must comply with each model’s license. Featherless provides infrastructure for third-party models and “does not license those models to you.” Featherless Terms of Service, Section 3, verified August 2, 2026
• Developer uses prepaid credits, not a separate bill. Your selected monthly credit amount is added to your balance each cycle; successful requests deduct from it by model price and token usage. Credits do not expire, and one-time top-ups are available. Featherless Request Pricing, verified August 2, 2026

Additional terms (arbitration, payments, liability)

Featherless processes payment via Stripe under Stripe’s own terms. Disputes go to binding arbitration under New York law, and Featherless disclaims liability for damages arising from use of the Services (Terms of Service, Sections 6 and 8). These clauses rarely affect a buying decision for a personal Hermes backend, but read them if you intend to ship something customer-facing.

Before you set up Hermes Agent (self-hosted route)

You need:

• a macOS or Linux terminal, or Windows using WSL2;
• a Featherless account on a Developer / token-based plan (256K context — required for Hermes);
• the Featherless API key from your account dashboard;
• a current Hermes Agent installation.

Direct links:

• Create a Featherless account
• Subscribe (Developer / token-based)
• Open the API keys dashboard
• Browse the model catalog
• Hermes Agent installation guide
• Hermes Custom Endpoint docs

Configure Featherless in Hermes as a Custom Endpoint

Featherless is not a built-in Hermes provider. It is wired in through Hermes’ Custom Endpoint path, which is the same path used for Ollama, vLLM, SGLang, llama.cpp, LM Studio, LiteLLM, and any other server that speaks the OpenAI Chat Completions API. Hermes Agent AI Providers, “Custom & Self-Hosted LLM Providers,” verified August 2, 2026

There are two equivalent routes: the interactive wizard and a named custom provider in config.yaml. Use the 256K Developer/token-based plan so you clear Hermes’s 64K minimum.

Option A: interactive wizard (recommended)

hermes model

In the prompts:

1. Choose Custom endpoint.

2. When asked for the API base URL, enter:

   https://api.featherless.ai/v1

3. When asked for the API key, paste your Featherless API key (created above).

4. When asked for the transport / API mode, choose OpenAI Chat Completions (chat_completions).

5. When asked for the model name, enter a slug from the Featherless catalog, for example Qwen/Qwen2.5-7B-Instruct, MiniMaxAI/MiniMax-M3, or moonshotai/Kimi-K3. Featherless Quickstart guide, verified August 2, 2026; Featherless model catalog, verified August 2, 2026

The wizard stores the provider configuration for you and writes it to ~/.hermes/config.yaml. Hermes Agent AI Providers, “General Setup,” verified August 2, 2026

Option B: named custom provider in config.yaml

To keep the key out of config.yaml, store it in ~/.hermes/.env and reference it via key_env:

mkdir -p ~/.hermes
chmod 700 ~/.hermes
read -rsp 'Featherless API key: ' FEATHERLESS_KEY; printf '\n'
printf 'FEATHERLESS_API_KEY=%s\n' "$FEATHERLESS_KEY" >> ~/.hermes/.env
unset FEATHERLESS_KEY
chmod 600 ~/.hermes/.env

Then add a named provider to ~/.hermes/config.yaml:

providers:
  featherless:
    api: https://api.featherless.ai/v1
    key_env: FEATHERLESS_API_KEY
    transport: chat_completions
    default_model: Qwen/Qwen2.5-7B-Instruct
    context_length: 256000   # Developer/token-based ceiling; clears Hermes's 64K minimum

key_env reads the secret from ~/.hermes/.env; transport: chat_completions matches Featherless’s OpenAI-compatible API; context_length: 256000 keeps Hermes above its required minimum. Hermes Agent AI Providers, “Named Custom Providers,” verified August 2, 2026

Inside an active Hermes session, /model featherless:Qwen/Qwen2.5-7B-Instruct switches to the Featherless endpoint mid-conversation. Because Featherless exposes an enormous catalog, always use the full org/model form — do not rely on bare /model custom, which is meant for servers exposing exactly one loaded model. Hermes Agent AI Providers, “Switching Models with /model,” verified August 2, 2026

If you select a Featherless model tagged Vision and want native image input through the custom provider, add supports_vision: true to that provider block; without it, Hermes may route the image through an auxiliary vision model instead of sending it natively. Hermes Agent AI Providers, verified August 2, 2026

Command map

The command syntax above follows the current Hermes CLI and Featherless quickstart. Because the endpoint is saved by hermes model, the smoke test does not need to repeat the base URL. Note: hermes chat does not take a --base-url flag — if you have not yet selected the provider, run hermes model first.

Smoke-test the connection

Run a small one-shot request. Because you configured the provider with hermes model, the simplest test is:

hermes chat -q "Reply with exactly: FEATHERLESS_OK"

A response containing FEATHERLESS_OK proves that Hermes can resolve the custom provider, read the credential, reach Featherless, and invoke the chosen model. It does not prove that every other model in the catalog works on your plan, that you have enough concurrent-unit headroom for a long-running agent, or that image or tool input is enabled.

Equivalent direct call (bypasses Hermes) if you want to sanity-check the key and endpoint first — make sure FEATHERLESS_API_KEY is exported in the current shell:

curl https://api.featherless.ai/v1/chat/completions \
  -H "Authorization: Bearer $FEATHERLESS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen2.5-7B-Instruct",
    "messages": [
      {"role": "user", "content": "Reply with exactly: FEATHERLESS_OK"}
    ]
  }'

The endpoint URL and the request shape come straight from Featherless’s quickstart. Featherless Quickstart guide, verified August 2, 2026

Troubleshooting

401 or “invalid API key”

Copy a newly created API key from the Featherless API Keys dashboard. Do not use an account password, billing identifier, or a key from another provider. Check for trailing whitespace; rerun hermes model. If you used Option B, confirm FEATHERLESS_API_KEY is exported in the shell that runs Hermes, or that ~/.hermes/.env is owner-readable. Featherless Quickstart guide, verified August 2, 2026

Hermes rejects the model for low context

You are on the Chat plan (32K) or set context_length below 64K. Hermes requires at least 64K. Move to the Developer / token-based plan (256K) and set context_length: 256000. Hermes Agent providers — “Minimum context: 64K tokens,” verified August 2, 2026

The key works but the request is rejected

The most common cause is an inactive plan. An API key can exist before a paid plan. Buy the plan, then retry with the same key. Featherless home, verified August 2, 2026

HTTP 429 (Too Many Requests)

You hit the concurrent-unit ceiling. With Chat’s 4 units (or the token-based plan’s budget), you can only run a fixed number of parallel requests by model size. Either wait for in-flight requests to finish, switch to a smaller model class, or upgrade concurrency. Featherless Concurrent Unit Limits, verified August 2, 2026

Hermes reports model not found for a valid slug

Featherless adds and removes models frequently; a slug may have changed or be unavailable on your plan. Recommended order: (1) copy the exact slug from the live catalog; (2) test it directly against /v1/chat/completions; (3) run hermes model --refresh; (4) reconfigure the endpoint if needed. Featherless model catalog, verified August 2, 2026

Tool calls come back as plain text

You picked a model that does not advertise tool calling in the catalog. Filter by the Tools tag in the catalog, or pick a known tool-capable slug such as Qwen/Qwen3-Coder-30B-A3B-Instruct or MiniMaxAI/MiniMax-M3. Featherless model catalog, “Tools” tag, verified August 2, 2026

Image input is silently dropped

Not every model on Featherless advertises Vision, and a custom provider may route images through an auxiliary vision model unless configured for native vision. Pick a Vision-tagged model and, for native input, add supports_vision: true to the provider block in config.yaml. Featherless model catalog, “Vision” tag, verified August 2, 2026; Hermes Agent AI Providers, verified August 2, 2026

Security and cost controls

Treat the Featherless API key like a paid credential. Keep ~/.hermes/.env owner-readable only, do not commit it, do not paste it into prompts, and rotate or replace it if exposed. Featherless processes payment via Stripe under Stripe’s own terms; the account holder is responsible for activity under the account. Featherless Terms of Service, Section 4, verified August 2, 2026

Two controls worth setting on day one:

• Match context_length to your plan. For self-hosted Hermes, set it to 256000 (Developer/token-based). Never set 32000 — that is below Hermes’s minimum and the agent will not start.
• Distinguish Chat from Developer/token-based. Chat is flat-rate with a 32K ceiling (not usable for Hermes). The token-based plan is prepaid credits; a runaway agent drains your balance, so monitor the billing dashboard and set a sensible monthly credit amount. If you intend to run unattended agents, prefer the token-based plan with explicit budget limits over Chat.

For a customer-facing application, scheduled batch system, or workload requiring stable throughput, do not stretch Chat or Developer beyond their stated use. The Terms of Service explicitly terminate individual subscriptions used for resale, batch, or multi-user production, with no refund. Featherless Terms of Service, Section 3, verified August 2, 2026

Keep this setup current

Featherless changes model coverage, plan shape, and pricing; Hermes changes provider wiring. Before buying or debugging, re-check:

1. the Featherless home page (current Chat/Developer pitch);
2. the Featherless Plans docs (full tier matrix);
3. the Featherless model catalog (exact slugs and Vision/Tools tags);
4. the Featherless Concurrent Unit Limits doc (per-model unit cost);
5. the Featherless Managed Hermes doc (marketplace launch);
6. the Featherless Request Pricing doc (prepaid credits and top-ups);
7. the Hermes Agent AI Providers doc (Custom Endpoint and named-provider syntax);
8. the Featherless Quickstart guide (current base URL and request shape).