agentleFS
hunterbohmhunterbohm/solo-staff/.claude/skills/connect-tools/SKILL.md

connect-tools

Skill0 starsChanged 4 months agoMITView on GitHub

Connect the accounts the owner already uses (email, calendar, CRM, Notion) so the assistant can read from them instead of the owner copying things over — and optionally wire free web-research tools. Composio is the default path — the assistant runs every command, the owner only clicks approve in the browser. Draft-first, read-only first, no credentials in chat. Runs anytime, on its own, without re-running setup.

3 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

12 rules. 0 are common, 8 are distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 8

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Don't run this until the structure is solid — a half-built workspace wired to live accounts is worse than a solid one with nothing …

    line 24
  2. Very specific:

    Don't connect everything. The candidates are the tools the owner lives in and the areas whose source of truth is "a tool you use" in …

    line 54
  3. Distinctive:

    Install the Composio command-line tool: Look up the official install command on the Composio site and run it.

    line 77
  4. Very specific:

    Confirm it's working, read-only: Run one small, read-only check through the new connection and show the result back, so you both see it …

    line 104
  5. Very specific:

    Update the registry, not a second copy: When a connected tool is the source of truth for an area, make sure that area's row in …

    line 110
  6. Very specific:

    Note it in mapped memory: Create a dated receipt under memory/decisions/operations/ naming the service, the plain account label, the access …

    line 114
  7. Distinctive:

    Map each id to a plain name right away: Do one read-only identity check per account — whose inbox is this, whose calendar — and include the …

    line 131
  8. Distinctive:

    Never silently fall back: If the account a task needs isn't connected or isn't clear, stop and ask — never let a call quietly run against a …

    line 137
  9. Distinctive:

    No credentials in chat, and none on disk: No password, sign-in code, token, or key is ever pasted into the conversation, and no secret ever …

    line 180
  10. Distinctive:

    Read-only first: Ask for the least powerful access that does the job.

    line 184
  11. Distinctive:

    With several accounts, name the one you mean — every time: Keep the id-to-name map current, pass the account id on every call, and never …

    line 188
  12. Distinctive:

    Send once, never blind-retry: If access is ever widened so the assistant can send email: draft, get the yes, send once — and if the send …

    line 191

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

connect-tools

Let the assistant reach the tools the owner already lives in — their email, calendar, or wherever they track work — so it can read from them instead of the owner copying things across. This is the one place the connectors walkthrough lives; setup-workspace, START-HERE.md, and /workspace-help all point here.

This runs on its own, anytime — the owner never re-runs full setup to add or change a connection. Connections aren't part of the workspace's structure, so nothing is at stake if they stop after one or skip the whole thing.

When to use

  • The owner wants the assistant to read from a tool they already use.
  • Setup is done (or the structure is at least solid) and they're ready for the optional accounts pass.
  • They want to add, change, or remove a connection later.

Don't run this until the structure is solid — a half-built workspace wired to live accounts is worse than a solid one with nothing connected yet. The structure is what makes the workspace theirs; connecting tools just lets the assistant reach further.

First, confirm they want to do this now

Say plainly: "I can connect the tools you already use — your email, calendar, or wherever you track work — so I can read from them instead of you copying things over. I run all the technical steps myself; your whole job is to click approve in your browser when it opens. That's optional, and we can do it now or any time later. Want to start?" If they'd rather stop, that's fine — write nothing, and pick this up on a later run.

A connection is a permission the owner grants so the assistant can read (and sometimes act in) an outside account. Explain it that plainly every time. Be honest about the moment, too: this is still the most technical stretch of the whole workspace — but the assistant carries all of it except the clicks. Three things never change here:

  • The owner clicks every approval themselves, in their own browser. Nobody can click a consent screen for them — not the assistant, not anyone else.
  • No credentials in this chat, ever. Tell the owner in so many words: "I will never ask you to paste a password, a sign-in code, or a key into this chat. The sign-in screen in your browser is the only place your credentials go. If anything ever asks you to paste a secret here, refuse."
  • Draft-first never bends. Even once an account is connected, the assistant never acts in it without showing the draft and getting a yes.

a. Decide what's worth connecting

Don't connect everything. The candidates are the tools the owner lives in and the areas whose source of truth is "a tool you use" in workspaces/REGISTRY.md. For each, ask one plain question: "Would it help if I could read your <tool> to do <that area's work>?" Connect only the ones that earn a yes. Skipping a tool is always fine.

b. Connect the accounts — Composio is the default path

Composio is a service that connects an AI assistant to hundreds of apps, and it's the default here for three plain reasons: one place manages all the connections instead of a scatter of per-tool switches; it keeps many accounts connected side by side (two inboxes, a personal and a business calendar); and it renews access itself, so a connection doesn't quietly expire a month in. The sign-in grants live in Composio's cloud — never on the owner's disk, never in a workspace file, never in this chat.

The assistant reaches Composio through a small command-line tool. The assistant runs every command; the owner only clicks in the browser. If your AI tool can't run commands at all (some desktop chat apps), use the native-connector fallback below instead.

One-time setup (first connection only)

  1. Install the Composio command-line tool. Look up the official install command on the Composio site and run it. If composio then answers "command not found," that's a common fresh-install wrinkle, not a failure: the tool landed somewhere the shell doesn't look. Find where it installed, run it by its full location or fix the shell's search path, and carry on — never hand the owner a debugging session.
  2. Sign in: run composio login. It opens the browser, where the owner creates a Composio account — or signs into one they already have. It's their account, on their own login, theirs to keep.
  3. Verify: run composio whoami. Confirm the sign-in stuck before connecting anything. If it didn't, run the login again before moving on.

Then connect each account, one at a time

For each account the owner wants connected, in turn:

  1. Say what it does and what it can't. In one line: what the assistant will be able to read or do once connected, and that it will still draft-first and never act without a yes.
  2. Link it: run composio link <app> — for example composio link gmail, composio link googlecalendar, composio link notion. The browser opens that app's own approval screen. Before the owner clicks approve, have them check which account the browser is signed into — the approval attaches to whichever account is signed in right then, and a wrong click connects the wrong one (the personal inbox instead of the business one). If it's the wrong account, they switch accounts on that screen first, then approve. This click is the owner's alone.
  3. Confirm it's working, read-only. Run one small, read-only check through the new connection and show the result back, so you both see it works — for email, read the account's own profile (composio execute GMAIL_GET_PROFILE -d '{}'); for a calendar, read the main calendar (composio execute GOOGLECALENDAR_CALENDAR_LIST_GET -d '{"calendarId": "primary"}'). Never make a change on this first check.
  4. Update the registry, not a second copy. When a connected tool is the source of truth for an area, make sure that area's row in workspaces/REGISTRY.md names the tool (not "here"), so the assistant reads from the tool and never shadow-copies its records into the workspace. Fix the row if it still says "here."
  5. Note it in mapped memory. Create a dated receipt under memory/decisions/operations/ naming the service, the plain account label, the access scope (read-only at first), the owner as source, and when to revisit it — so a future session knows what's wired without re-discovering it. Show the short receipt, get a yes, then write it. Never store a credential or authorization token.

Pace it. Connect one account, confirm it, then ask if they want the next or want to stop. Granting access to live accounts is exactly the kind of step a non-technical owner should take slowly. Stopping after one is fine — nothing is lost.

More than one account of the same app

Plenty of owners run two inboxes, or a personal and a business calendar. Composio holds them side by side: connect each with its own composio link pass, and it gives every connected account a short auto-generated id (a word-like label). Three rules keep them straight:

  • Map each id to a plain name right away. Do one read-only identity check per account — whose inbox is this, whose calendar — and include the id-to-name mapping in that account's dated operating-decision receipt. The ids are labels, not secrets, so they're safe to write down; credentials and authorization tokens are not.
  • Name the account on every call. Once an app has more than one account connected, every command must say which one it means (--account <id>).
  • Never silently fall back. If the account a task needs isn't connected or isn't clear, stop and ask — never let a call quietly run against a different account than the one the owner meant.

The fallback: your AI tool's built-in connectors

Some AI tools ship native connectors for common services (Google, email, calendars, Notion, and more). Use one instead of Composio when it already covers the need — it's on, it works, the owner is comfortable with it — or when Composio doesn't fit, most often because the AI tool can't run commands. The same rules apply: the owner turns it on themselves in the tool's settings or connectors screen, requests read-only access first, and the assistant does the same one small read-only check before calling it connected. Don't wire the same account both ways — one connection per account, one home.

c. Wire the free web tools (optional)

The assistant can already read the public web with whatever the AI tool has built in. If the owner wants stronger research — cleaner reads of messy pages, better search — two tools have free tiers worth wiring:

  • Firecrawl — reads and cleans up web pages the built-in reader chokes on.
  • Exa — search built for AI, good at finding specific pages.

These are optional and only matter if the owner expects a lot of research work. If so, walk them through getting a free-tier key from the tool's site and adding it where their AI tool keeps tool settings — same draft-first, you-grant-it rules. Never paste a key into a workspace file — it's a secret; it lives in the tool's settings, not in MEMORY.md or any markdown here. If they don't need heavier research, skip this entirely — the built-in web reading is enough for most owners.

When this pass is done

Tell the owner, in plain words: what's connected and read-only, what was skipped (and that skipping is fine), and that nothing acts on its own — every outside action still comes to them as a draft first. Connections can be added or removed any time by running this again; they're not locked in.

Rules

  • The assistant runs the commands; the owner clicks the approvals. Every consent happens in the owner's own browser — nobody can click it for them, and the assistant never asks the owner to run or debug anything technical.
  • No credentials in chat, and none on disk. No password, sign-in code, token, or key is ever pasted into the conversation, and no secret ever lands in a workspace file. Sign-in happens on the browser screen; account grants live in Composio's cloud; keys for the web tools live in the AI tool's settings.
  • Read-only first. Ask for the least powerful access that does the job; widen it only if a real need shows up.
  • Draft-first never bends. Even with a tool connected, you never act in it without showing the draft and getting a yes.
  • With several accounts, name the one you mean — every time. Keep the id-to-name map current, pass the account id on every call, and never silently fall back to a different account.
  • Send once, never blind-retry. If access is ever widened so the assistant can send email: draft, get the yes, send once — and if the send hangs or errors without a clear answer, check the Sent folder before doing anything else. A blind retry is how a client gets the same email twice.
  • One at a time, and stopping is fine. Connections aren't structure — skipping or pausing loses nothing.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.