Hermes Agent Not Responding on Telegram: 4 Fixes

A Hermes Agent that goes silent on Telegram has four documented causes, and they announce themselves differently: a gateway that stopped, an authorization gap, a model that the agent refuses to use, or a model that no longer exists at the provider. The first three come from the official Hermes Agent FAQ and the project's issue tracker during 2026. The fourth comes from our own hosting fleet, measured in August 2026. This page gives the check for each, in the order that finds the culprit fastest.

Check 1: is the gateway even running?

The gateway is the process that holds the Telegram connection open. The official FAQ lists three states behind a silent bot: the process stopped, the authorization failed, or the sender is missing from the allowlist. All three checks start with two commands on the server:

hermes gateway status
cat ~/.hermes/logs/gateway.log | tail -50

What this means for you: a stopped gateway is the boring answer and the most common one. If status says it is down, start it with hermes gateway start, watch the log, and send one test message. Everything below assumes the process is up.

Check 2: authorization and the allowlist

Hermes decides who it answers. If your Telegram user is outside the configured allowlist, the agent receives the message and stays silent by design. The FAQ points to hermes config show for verifying the bot token and the authorization mode. Three settings matter: whether the channel uses an allowlist, DM pairing, or open access, and whether your numeric Telegram ID is actually on the list.

What this means for you: if the bot answers one person and ignores another, this is almost certainly the layer. It reads as a malfunction and it is a permission.

Check 3: the 64,000-token floor

On May 12, 2026, issue 24140 reported a Telegram bot that went down for every user overnight. The log carried the same line each time:

Model MiniMax-M2.7 has a context window of 32,768 tokens, which is below
the minimum 64,000 required by Hermes Agent. Choose a model with at least
64K context, or set model.context_length in config.yaml to override.

Hermes Agent enforces a context floor: the working memory a model brings has to fit the agent's system prompt, its tools and your conversation. A model with a 32K window fails that test, and the agent rejects it for every message rather quickly. The same report shows the identical error firing on a scheduled job with a second 32K model, so the floor applies everywhere the model is used, chats and crons alike.

What this means for you: if the silence started right after you changed models, or your provider swapped a model version under the same name, read the log for the words below the minimum. The fix is one of two moves, both named in the error itself: pick a model with a 64K or larger context window, or set model.context_length in the config if you know the model's real window is larger than what was detected. The issue was closed on May 17, 2026, five days after it was opened.

Check 4: the model that no longer exists

A provider can retire a model while your config still names it. In August 2026 we measured this failure mode on our own Hermes hosting fleet, and its signature is specific: gateway up, authorization fine, and every model call failing, because the configured model had been withdrawn by the provider in the meantime. Since August 28, 2026 our provisioning verifies the configured model against the provider's live list and stores the result, so an agent starts on a model that is confirmed to exist.

The check on any server takes one minute: run hermes model, which the FAQ documents as the place to add providers and enter keys, and confirm the model your config names still appears on the provider's current list. A model that vanished upstream produces a bot that looks healthy everywhere except the one log line that matters.

The imposter: a typing indicator that never stops

One Telegram symptom looks like a hang and is only cosmetics. Issue 28004, opened May 18, 2026, documented the typing indicator staying visible indefinitely after the reply had already been delivered. The cause was a race in the code that refreshes the indicator every two seconds: one last refresh could fire after the reply went out, and Telegram's Bot API has no way to cancel a typing signal, so it lingers until it expires on its own. The issue was closed on July 13, 2026.

What this means for you: if the answer arrived and the bot still shows typing, nothing is stuck. Update to a current build and the ghost typing goes away.

The five-minute order of operations

1. Process: hermes gateway status. Down means start it and retest.
2. Log: tail -50 the gateway log. Read the newest error line before anything else.
3. Permission: hermes config show. Confirm the token and that your Telegram ID is allowed.
4. Model floor: search the log for below the minimum. Swap to a 64K+ model if found.
5. Model existence: hermes model. Confirm the configured model is still offered by the provider.

An agent that answers at 3 a.m. needs a machine that is awake at 3 a.m. and a model that still exists at 3 a.m. Our managed Hermes Agent hosting keeps the gateway monitored and verifies the model against the provider's live list before your agent ever says hello, with a 7-day free trial and no card required.

Sources: the official Hermes Agent FAQ, Hermes Agent issues 24140 and 28004, and measurements on our own hosting fleet dated August 25 to 28, 2026.