Agent Skills › memoe-app/memoe › composio-router

composio-router

GitHub

指导如何高效调用Composio的500+工具,重点在于控制结果体积、理解参数默认值及路径差异,避免产生过大响应导致对话中断。

moe-home/hermes-skills/connectors/composio-router/SKILL.md memoe-app/memoe

Trigger Scenarios

需要调用Composio工具时 处理Gmail/Slack等第三方API集成时

Install

npx skills add memoe-app/memoe --skill composio-router -g -y
More Options

Non-standard path

npx skills add https://github.com/memoe-app/memoe/tree/main/moe-home/hermes-skills/connectors/composio-router -g -y

Use without installing

npx skills use memoe-app/memoe@composio-router

指定 Agent (Claude Code)

npx skills add memoe-app/memoe --skill composio-router -a claude-code -g -y

安装 repo 全部 skill

npx skills add memoe-app/memoe --all -g -y

预览 repo 内 skill

npx skills add memoe-app/memoe --list

SKILL.md

Frontmatter
{
    "name": "composio-router",
    "author": "Moe",
    "license": "MIT",
    "version": "1.0.0",
    "metadata": {
        "hermes": {
            "tags": [
                "Composio",
                "MCP",
                "Connectors",
                "Google",
                "Gmail",
                "Cost"
            ],
            "toolkit": "*",
            "related_skills": [
                "composio-gmail",
                "composio-google-calendar",
                "composio-google-drive",
                "composio-google-docs",
                "composio-google-sheets",
                "composio-google-tasks",
                "composio-google-meet"
            ]
        }
    },
    "description": "Read FIRST before any Composio tool call — Gmail, Calendar, Drive, Slack, Notion, GitHub, any of the ~500 toolkits. How to pick the tool, how to call it without dumping a megabyte into the turn, and the four things Composio's own schemas will not tell you. Also: how to check whether a service is connected WITHOUT triggering a connection request."
}

Composio, and how to call it without paying for it twice

Composio is one MCP session standing in front of ~500 toolkits. Its tools are authored one at a time, by different hands, so there is no convention to learn: two parameters with the same name in the same toolkit mean different things, defaults are as often the expensive setting as the cheap one, and a few advertised parameters do nothing at all.

This skill is the shape. It is true of toolkits nobody has written a skill for. Read it before the first call, not after the result comes back too big.

Why size is the whole game

An oversized tool result is not delivered to you. The client writes it to a file and hands you a path — so answering "who emailed me" turns into grepping a file for an answer a cheaper call would have put straight in front of you. The cost is not the bytes; it is that the answer stops being in the conversation. A result of a few megabytes from one Gmail call is ordinary.

So the discipline is not "be tidy". It is: the cheap call is the one that answers the question, and the expensive call is the one that does not.

The four questions, before any Composio call

Composio's schemas answer none of these. Only a measurement does.

1. What does this parameter default to? Assume the expensive setting until you have checked. Measured on Gmail: verbose and include_payload both default to the costly value, and the two are independent — same five messages, GMAIL_FETCH_EMAILS called directly:

both defaults            167,231 B
include_payload: false    49,071 B     3.4x
verbose: false            37,064 B     4.5x
both false                 4,979 B    34x   (995 B/message)

Nothing in either description mentions bytes or tokens. verbose describes speed ("~75% improvement") and then content; a model optimising for size gets no signal from it at all.

2. Does a parameter that talks about speed actually control size? Yes, on GMAIL_FETCH_EMAILS. The name is not the guide and neither is the description's subject — the only guide is the byte count.

3. Is the parameter honoured at all? GOOGLECALENDAR_EVENTS_LIST_ALL_CALENDARS.max_results_per_calendar is documented as "maximum events returned per calendar" and is inert — 1, 2 and omitted are byte-identical, verified again 2026-09-04. A cap you believe in and that does nothing is worse than no cap, because you stop looking.

4. Does the invocation path change the answer? Sometimes, and it is per-tool. Same query, same five messages, same day:

GMAIL_FETCH_EMAILS   direct 167,231 B   |   via MULTI_EXECUTE  49,218 B
    and through the multiplexer `include_payload` is INERT
    (49,218 B with and without, byte-identical) — `verbose` is the
    only lever there (5,126 B).

GMAIL_FETCH_MESSAGE_BY_MESSAGE_ID   direct 54,837 B | muxed 54,999 B
    identical but for a 162-byte envelope. Path-insensitive.

So "the path changes the answer" is a property of the tool, not of Composio. Do not carry a lever from one tool to the next; check it on the tool you are about to call.

A wrong value costs more than no value

GOOGLEDRIVE_FIND_FILE.fields takes a Google field selector. Measured:

omitted                    1,041 B     (the documented light default)
fields: "files(id,name)"     555 B     a real selector — cheapest
fields: "basic"            4,008 B     not a selector
fields: "*"                4,006 B     everything

An unrecognised selector does not error. It falls back to everything — four times the cost of omitting the parameter you were trying to economise with. A well-formed value that resolves to nothing is more dangerous than an absent one, because absence is honest.

Some inputs are outputs

GOOGLEDRIVE_FIND_FILE advertises 22 input properties and nine of them are things the tool tells you about what it did to your request — six booleans described "Indicates whether …" (pagetoken_dropped, email_query_transformed, editors_field_removed, …) and three original_* strings. GOOGLESHEETS_BATCH_GET has empty_strings_filtered. They are harmless to pass and they are not free: they are 40% of that tool's 11.9 KB schema, and they invite you to set a flag that means nothing.

Pick the TOOL first. No parameter fixes the wrong one.

To answer "who emailed me today" you never need the by-id fetch. A list tool answers it; the by-id fetch answers "what does this one message say", and costs 55 KB at its default to do it. Every toolkit has this split, and the names do not always make it obvious — GOOGLEDOCS_GET_DOCUMENT_PLAINTEXT exists beside GOOGLEDOCS_GET_DOCUMENT_BY_ID and answers a different question far more cheaply.

Ask for one before you ask for a hundred. One item, look at the size, then scale. That is the whole method, and it is the only one that survives a toolkit nobody has written down.

Calling: the three paths

  1. Preloaded tools are in your tool list already — a handful of reads for the connected toolkits, plus the five COMPOSIO_* meta tools. (One real session, measured: 13 tools, of which 8 were Google reads, and tools/list cost 52,608 B to deliver them — paid on every child spawn.) Call these by name.
  2. Anything else — found by search — is not individually callable. Run it through COMPOSIO_MULTI_EXECUTE_TOOL with {"tools": [{"tool_slug": "…", "arguments": {…}}]}. Remember question 4.
  3. Discovery is COMPOSIO_SEARCH_TOOLS with {"queries": [{"use_case": "<plain English>"}], "session": {"generate_id": true}}, then COMPOSIO_GET_TOOL_SCHEMAS with tool_slugs: [...] for the exact parameters. Do not guess slugs.

Discovery is not free. Measured 2026-09-04: one COMPOSIO_SEARCH_TOOLS query costs 7–27 KB depending on the toolkit; COMPOSIO_GET_TOOL_SCHEMAS costs 18–36 KB for twelve slugs. The search tool's own schema encourages you to split work into more queries — each split is another 7–27 KB. Ask one good question.

Connection status — the one meta tool with a side effect

COMPOSIO_MANAGE_CONNECTIONS is not a status check. It MINTS a connection request for anything unconnected, so using it to find out whether Gmail is connected creates the state you were asking about.

Calling a tool for an unconnected toolkit returns a ~420–460 byte refusal that names the toolkit and tells you to call COMPOSIO_MANAGE_CONNECTIONS. Do not follow that instruction on your own initiative. Tell the person the service is not connected and let them ask for it.

The pure read is COMPOSIO_SEARCH_TOOLS, whose result carries toolkit_connection_statuses — per toolkit, has_active_connection and, where connected, the account it is on.

Schemas are readable for toolkits that are not connected. A search against an unconnected toolkit returns real slugs and real parameter definitions. So you can always find out what a call would take; what you cannot know without a connection is what it would cost.

Never write an identifier down

Composio's user_id is a capability, not a name: anything holding it can act as that person. It must never appear in a message, a note, a commit, a file you write, or a summary you hand back. Fold it — abcd…yz — if you have to refer to it at all.

The per-toolkit skills

composio-gmail, composio-google-calendar, composio-google-drive, composio-google-docs, composio-google-sheets, composio-google-tasks, composio-google-meet.

Each exists because that toolkit contradicts something above — an expensive default, an inert parameter, or a tool choice the names hide. Google Photos and Google Slides have no skill on purpose: everything true of them is already on this page (Photos' page sizes default to a sensible 20–25; Slides' fields is one more instance of the selector rule). An empty skill that exists is worse than an absent one, because opening it costs a call.

For any of the other ~490 toolkits, this page IS the skill. Ask the four questions.

Version History

  • e2735d6 Current 2026-11-20 21:42

Same Skill Collection

moe-home/hermes-skills/connectors/composio-gmail/SKILL.md
moe-home/hermes-skills/connectors/composio-google-calendar/SKILL.md
moe-home/hermes-skills/connectors/composio-google-docs/SKILL.md
moe-home/hermes-skills/connectors/composio-google-drive/SKILL.md
moe-home/hermes-skills/connectors/composio-google-meet/SKILL.md
moe-home/hermes-skills/connectors/composio-google-sheets/SKILL.md
moe-home/hermes-skills/connectors/composio-google-tasks/SKILL.md
moe-home/hermes-skills/writing/humanizer/SKILL.md

Metadata

Files
0
Version
e2735d6
Hash
2bffd462
Indexed
2026-11-20 21:42

Accueil - Wiki
Copyright © 2011-2026 iteam. Current version is 2.156.0. UTC+08:00, 2026-11-20 23:49
浙ICP备14020137号-1