Five names in the Hermes platform registry reported the same tool schema on an empty home. cli, discord, telegram, cron, and slack each exited 0 from hermes prompt-size --json with 19 tools and 35850 bytes of tool-schema JSON. Their system prompts did not match. The CLI prompt was 6579 bytes. Discord was 6198. Telegram was 6326. Cron was 6591. Slack was 6198, the same prompt size as Discord and a different stable tier by 2 bytes.
That split is the number you want before you disable a toolset to shrink a gateway. The tool count is the platform check. The prompt size is the hint. A second group of names also exited 0, and those names reported 3 tools and 2287 schema bytes: tui, desktop, the typo nope, and the uppercase forms CLI and Discord. Exit status does not separate the groups.
What this measurement used
The binary was Hermes Agent v0.22.5+471.g5894e75, release line 2026.9.24, on Linux x86_64, on 21 November 2026. hermes --version printed that version line. Each prompt-size invocation used a new Hermes home, created for the run, with no config.yaml, no installed skills, and no memory files before the first command. The operator home was not the home under test. No login was run. No prompt was sent to a model. The command’s own reference says it builds the system prompt and inspects tool schemas offline. The model field in every JSON report was empty, and stderr was empty.
The help text for --platform says the value simulates a platform, with cli, telegram, and discord as the examples, and that the default is cli. It does not list a closed set, and the parser does not reject a string that is absent from the registry. The public CLI reference at https://hermes-agent.nousresearch.com/docs/reference/cli-commands describes the same command as an offline byte breakdown of the system prompt and tool schemas. It does not publish these counts. The version-matched page shipped with this binary adds the --platform examples and the telegram simulation sentence. It does not list the registry, and it does not say an absent name still exits 0.
The JSON fields that matter are tools.count, tools.json_bytes, and system_prompt.bytes. skills_index, memory, and user_profile were 0 bytes on this home. The skills index is the block the reference says is often the largest when many skills are installed. That sentence is about a populated home. It is not this one. On this home the tool schemas were the large block: 35850 bytes against a 6579-byte CLI prompt.
Five registry names, one tool schema
The five matching rows are registry names. The platform table in this tree maps cli to hermes-cli, discord to hermes-discord, telegram to hermes-telegram, cron to hermes-cron, and slack to hermes-slack. On the empty home those composites resolved to the same enabled set. The JSON toolset breakdown was the same list for each of them: file, skills, delegation, browser-use, memory, code_execution, a group labeled (unknown), terminal, web, clarify, and tts.
The JSON report does not print tool names. The names below were read from the inspection agent that prompt-size builds, on the same home, after the CLI runs. For cli and discord that list was the same 19 names, in the same order: browser_exec, clarify, delegate_task, execute_code, memory, patch, read_file, search_files, skill_manage, skill_view, skills_list, terminal, text_to_speech, web_extract, web_search, write_file, tool_search, tool_describe, and tool_call. Telegram, cron, and slack were not name-compared. Their JSON tool count and schema byte count matched the CLI row. That is the check a script can make without importing the agent.
The prompt bytes moved because the stable tier moved. The CLI stable tier was 5680 bytes. Discord was 5295. Telegram was 5422. Cron was 5691. Slack was 5297. The volatile tier, which the breakdown labels as memory, profile, and timestamp, differed from the CLI’s 897 bytes by the length of the platform name: Discord was 4 bytes longer, Telegram 5, cron 1. Memory and the user profile were empty, so that small gap is not a second copy of MEMORY.md. It is not the reason to pick one platform’s report over another.
The reference says --platform telegram simulates a messaging platform’s prompt, and that the difference is the platform hint. This run agrees with that sentence for the five names above, and adds the measurement the page does not: on an empty home the hint changes the prompt and leaves the tool schema where it was. A job that trims Discord by copying the CLI prompt size is using the wrong prompt and the right tool schema. A job that trims Discord by copying only the CLI tool count is using the right tool schema and still has not measured the Discord hint. Run the flag with the platform you will actually start.
Two smaller registry rows are not typos
webhook and api_server are in the same registry. They did not join the 19-tool row, and they should not. Treating every smaller count as a bad name will make you “fix” a platform whose default toolset is supposed to be smaller.
webhook exited 0 with 6 tools and 6000 schema bytes. The system prompt was 4214 bytes, and the stable tier was 3311, the same stable size as the typo row below. The tool names from the inspection agent were clarify, web_extract, web_search, tool_search, tool_describe, and tool_call. There is no terminal, no read_file, and no write_file. The toolset table in this tree defines hermes-webhook as a webhook toolset with a safe-tool list. The small prompt does not mean the name was missed. The tool count of 6, against that known list, means the registry row was found.
api_server exited 0 with 17 tools and 32356 schema bytes. The prompt was 6427 bytes. Compared with the CLI name list, the missing names were clarify and text_to_speech. The toolset breakdown omitted clarify and tts and kept file, skills, delegation, browser-use, memory, code_execution, (unknown), terminal, and web. That is a real registry default, hermes-api-server, not a fallback. If you disable tools from the CLI report and then apply that list to the API server, you will be editing a set this flag did not report for api_server.
A script that only checks exit status will accept all of these. A script that only checks that the prompt is non-zero will also accept them. The useful comparison is tool count and schema bytes against a row you have already recorded for that exact name.
Names that keep a hint and lose the tools
tui and desktop are not in the platform registry this flag consults. They are in the prompt builder’s hint table. The hint table has entries for cron, cli, tui, desktop, and api_server. It has no discord or telegram entry; those hints come from the platform adapters, which is why the measured Discord and Telegram prompts still differed from the CLI prompt even though the hint dict does not name them.
The tool resolver does not use that hint table. If the name is absent from the platform registry, the default toolset becomes hermes- plus the name you typed. This tree defines hermes-cli, hermes-cron, and hermes-webhook. It does not define hermes-tui or hermes-desktop. The measured result for both names was the same tool row as a typo.
tui exited 0 with a 4895-byte prompt, a 3996-byte stable tier, 3 tools, and 2287 schema bytes. desktop exited 0 with a 5636-byte prompt, a 4733-byte stable tier, and the same 3 tools and 2287 schema bytes. The inspection agent named those tools tool_search, tool_describe, and tool_call. The JSON breakdown does not print those names. It prints one toolset, (unknown), with a count of 3. A larger prompt is the hint still being attached. It is not evidence that the TUI or desktop session’s tools were loaded.
The typo nope landed on that same tool row: exit 0, empty stderr, 4211-byte prompt, 3311-byte stable tier, 3 tools, 2287 schema bytes, names tool_search, tool_describe, and tool_call. The stable tier matched webhook, and the tools did not. That is why a prompt-size comparison has to read both fields. An empty-looking prompt can be a real webhook row or a missed name. The tool count separates them on this binary.
Case is not folded. CLI exited 0 with 3 tools, 2287 schema bytes, and a 5087-byte prompt. Discord exited 0 with the same 3 tools and 2287 schema bytes, and a 4706-byte prompt. Neither matched the lowercase registry row. A job that title-cases the platform, or copies CLI from a heading, will measure the fallback tools and a prompt that is not the CLI hint. The help text’s examples are lowercase. The registry keys are lowercase. Pass them that way.
The three fallback tools are still real tool schemas. They are the deferred-tool catalog entry points, grouped as (unknown) because they are not in the static toolset map the breakdown uses. A count of 3 is therefore not an empty report. It is a specific small set. If your recorded CLI row says 19 and 35850, a later run that says 3 and 2287 did not measure that row, whatever the prompt size is and whatever the exit code is.
Verification
Create a home for the check. Do not point it at a profile you use for live sessions. The first prompt-size run will create directories under that home. That is the command’s own working state, not a model call.
mkdir -p "$HOME/prompt-size-lab"
export HERMES_HOME="$HOME/prompt-size-lab"
hermes prompt-size --platform cli --json
hermes prompt-size --platform discord --json
hermes prompt-size --platform webhook --json
hermes prompt-size --platform tui --json
hermes prompt-size --platform nope --json
On this binary and this kind of empty home, require exit 0 and empty stderr for each of those, then require the JSON fields above: 19 tools and 35850 schema bytes for cli and discord, 6 tools and 6000 schema bytes for webhook, and 3 tools and 2287 schema bytes for both tui and nope. The prompt byte counts in the opening are what this home printed. A later Hermes build can change them. The relation is the part to keep: the registry names that share a composite class share a tool count, the webhook row is smaller on purpose, and tui shares its tool count with a typo.
If you already have a saved platform_toolsets list, installed skills, or a disabled-toolset entry, do not expect these counts. The resolver uses an explicit saved list when one exists, and the skills index grows when skills are installed. Re-run the flag against that home, with the exact platform name, and compare the new JSON with a fresh empty-home run of the same binary if you need to see what the configuration added. Do not paste an empty-home CLI count into a Discord gateway that has its own saved list.
Do not use --platform tui or --platform desktop as the measurement of a TUI or desktop session. Those session kinds have prompt hints. This flag’s tool registry does not contain those names, so the tool half of the report is the fallback set. Measure cli for the classic terminal, and measure the messaging platform by its lowercase registry key before you disable anything from hermes tools.
This run did not start a session, did not record a billed token count, and did not compare the inspection agent with a live TUI turn. The offline report is the fixed budget the command is documented to build. It is not a receipt from a provider. If the number you needed was the provider’s prompt tokens, this flag is the wrong instrument, and a successful exit does not supply that receipt.