Teams Are Here: A Shared Swarm, a Shared Wallet, and Deny Rules That Always Win

BotTrottr Team · · Product

Until now BotTrottr was a one-person tool: your agents, your keys, your wallet. But a lot of you told us the same thing — “I’d love my teammate’s agents to work alongside mine.” Today that’s possible.

The Team plan is $69 a month for up to five people, and it comes with a shared swarm, a shared wallet, shared tool credentials, and guardrails the team admin controls. This post walks through what it does and — more interesting, we think — why it works the way it does.

What you get

  • A Team Swarm. Every team gets a default swarm (it can’t be deleted). New agents created by members land in it. Everyone’s agents in the swarm share one files folder, the swarm’s skills, and any MCP credentials the admin saves for the team.
  • A delegated wallet. The admin designates one member’s x402 wallet as the swarm’s wallet. Agents pay from it — and that member approves every spend.
  • Deny rules that win. Block actions outright, for the whole team, with rules members can’t turn off.
  • Full visibility. Everyone on the team can see all agents and their run history.
  • Verified invites. Invitations only work for the exact email address they were sent to — and only once that address is verified.

Deny rules: because “always ask” isn’t enough

BotTrottr already had a human-in-the-loop gate: when an agent wants to send an email or make a payment, it proposes the action and waits for you. You could also write auto-approve rules for the boring, safe cases.

The thing we were missing was the opposite: a way to say “never”.

If an agent can be tricked — a poisoned web page, a hostile email it was asked to summarize — it can propose something you’d never approve. An approval queue is a decent safety net, but a queue full of bad proposals wears people down. A deny rule stops the action before it reaches the queue, so there’s nothing to click through by accident.

A deny rule uses the same condition engine as auto-approve rules. A scope (provider, action type, optionally a swarm) plus a test on the action’s fields. For example:

Provider: Microsoft 365 · Action: send · Field: To (recipient) · Operator: has an address NOT ending in · Value: @mycompany.com

Now no agent can email outside your company, no matter what it’s told. A few details we sweated:

  • Deny always wins. It beats your own auto-approve rules and an agent’s “full auto-approve” switch.
  • Nothing is queued. The agent gets a clear “Permission denied” that names the rule and tells it not to retry.
  • Approval can’t override it. We re-check deny rules at the moment an approved action is about to run. A rule you create after something was queued still blocks it.
  • Recipient lists are handled properly. The email allowlist checks every address — in To, comma lists and Name <addr> forms — and the @ in the suffix means look-alike domains like evilmycompany.com don’t sneak through.
  • A typo can’t silently disable a rule. Conditions are picked from a dropdown (To, Subject, URL, Amount…), and the server rejects fields that don’t exist — a mistyped field would otherwise never match and a “deny” would quietly do nothing.

Spending caps that look at the real price

Payments were the interesting case. When an agent proposes an x402 payment it only knows the URL — the price isn’t known until the endpoint replies with its quote. So for payments we check deny rules a second time, after the quote and before anything is signed. A rule like Amount (USDC) > 0.05 or Payee is none of [your approved addresses] is evaluated against what the endpoint actually asked for. If it matches, nothing is paid.

(The wallet’s built-in per-payment and rolling limits still apply underneath. Deny rules only ever make things stricter.)

Four layers, one rule: deny wins

Rules now exist at several levels:

  1. Platform rules — set by BotTrottr admins, apply to everyone.
  2. Team rules — set by the team admin, apply to every agent in the team swarm.
  3. Your deny rules — for all your agents, or just one swarm.
  4. Your auto-approve rules — skip the human for matching actions.

Any deny, at any level, blocks the action. Team and platform rules are deny-only on purpose: an admin should never be able to auto-approve a send or a payment out of someone else’s account. Members see team and platform rules as a read-only list on their Approvals page, so a block never looks like a bug.

Sharing a wallet without sharing the keys

The wallet design is the part we’re proudest of, because the obvious version is the dangerous one.

The obvious version: “the team has a wallet; any agent can spend from it.” That means any member — or any prompt-injected agent — can spend money that belongs to someone else, approved by themselves.

So instead:

  • The admin picks one member as the wallet holder. The wallet stays that person’s own non-custodial wallet.
  • When a teammate’s agent proposes a payment, it shows up in the holder’s Approvals — and in their inbox. The person who asked can’t approve their own spend.
  • Auto-approve never applies to these spends. Not your rules, not the agent’s full-auto flag. A human who owns the money decides.
  • The payment is made from the holder’s wallet — the one that was designated when the proposal was created.
  • If the holder leaves the team, the swarm falls back to everyone’s own wallets automatically.

Add a team deny rule on amount or payee and you have a shared treasury with a spending policy.

One shared folder, shared skills, shared credentials

A personal swarm’s shared folder has always been private to its owner — a handoff area between your agents. The Team Swarm’s folder is different: one location visible to every member’s agents, so a research agent on one person’s account can drop a file that a writing agent on another’s picks up.

Skills work the same way. Upload a skill to the swarm and everyone’s agents can use it. And for tools: the admin enters credentials once — say, your team’s Jira or GitHub token — and agents in the swarm use them whenever they don’t have their own. Members never see the secret; they just see their agent can use the tool.

“Does leaving the team really cut me off?”

Yes — and not just on paper. A swarm is only a pointer on an agent, so we don’t trust the pointer. Every single run re-verifies that you still belong to the swarm before handing over its files, skills, credentials or wallet. Remove a member and their agents lose access on the very next run, even though the agent still says “I’m in the team swarm”.

While building this we also audited every path that touches swarm data and tightened the checks, so a swarm’s skills and files can only be reached by its owner or members of its team. If you’re building anything multi-tenant, that audit is worth doing before you add sharing, not after — sharing multiplies the cost of any gap.

Invites you can’t forward

Team invites are tied to an email address. To accept one, the account you sign in with must have that exact email, verified with your sign-in provider. The link alone isn’t enough, so forwarding the email doesn’t let someone else in, and if we can’t verify who you are we fail closed rather than let you in. Invites are single-use, expire after seven days, and an account can belong to one team.

Pricing, and what we decided about visibility

  • $69/month covers up to 5 members including the admin. Usage on platform models is pooled and billed to the admin at provider cost plus 5%, after a pooled $25 monthly credit. Members on their own keys (BYOK) aren’t metered.
  • Everyone sees everyone’s agents and run history. We debated private-by-default and chose transparency: it’s how teams actually debug and learn from each other’s agents. (It’s read-only — you can read a teammate’s runs, not edit or run their agents.)

Try it

Open Billing or the new Team page and choose Start Team plan. Your team and its swarm are created automatically, and you can invite your first teammate in under a minute.

The full setup guide is in the docs: Teams documentation. If you try deny rules on your own account first, the Approvals page has everything you need — the Deny switch is in the New rule form.

We’d love to hear what you build with it — and which rule you wish you could write that you can’t yet. Tell us at [email protected].

← Back to all posts