The shift from "does the connector exist?" to "what is the agent allowed to do?" is the whole game now. A valid token proves identity, but it doesn't stop a prompt-injected email from triggering a destructive action. Scoping permissions at the MCP server level per agent role is a much stronger boundary than just hoping the model behaves.
Exactly. The token tells you who the user is, not what the agent should be doing right now. Scoping per role means a prompt-injected "delete this" has nothing to call, so you're not relying on the model saying no.
exactly. relying on the model to "do the right thing" when it gets prompt-injected is a losing battle. if the token literally doesn't have the delete scope, it doesn't matter how clever the malicious prompt is. it's about moving security from the prompt level down to the infrastructure level.
I research AI tools, APIs & agents — comparing specs, pricing and real user reports — then publish honest rankings so you don't waste money. Research-first, never sponsored fluff.
The migration footnote is the bit I'd underline — refresh tokens tied to the issuing OAuth client means switching vendors later isn't a config change, it's re-authenticating every user. That's the kind of cost that never shows up in the catalog comparison and bites hardest exactly when you're big enough to care.
Yep, that's the one people find out about too late. Moving the config takes an afternoon. Asking every user to reconnect is a support project. Worth thinking about who owns the OAuth apps from day one.
Good framing — the real production gate is authorization and auditability, not connector count: per-tenant token isolation, least-privilege scoping, and an audit trail that ties each action to a user and permission are what the security questionnaire actually probes. The same split shows up in local/desktop agents too: stored credentials still need scoped isolation and an action log, and a self-hostable or VPC-only option becomes a hard requirement once data can't leave the boundary. Practically, that means picking the big managed catalog for fast demos, but leaning toward code-you-own, self-hostable integration when compliance and custom auth (RSA keys, domain-wide delegation) matter.
Agreed, and good point on local and desktop agents. Same problem, smaller blast radius, but the scoping and the action log still matter. Big catalog for the demo, then re-check once compliance shows up. That's pretty much the path most teams take.
I write about Next.js + TypeScript + AI engineering, with a focus on making sites discoverable to LLMs and AI search. Long-form notes at mudassirkhan.me
the question changing from does a connector exist to what can the agent do once connected is the whole story imo. the one that bit us was your number 2. we gave an agent one token per customer and it could read every mailbox that token reached, which was not what the user actually approved.
per call logs tied to a user and an agent id were what finally made the security review boring.
how do the three handle revoking one tool mid session without killing the whole connected account?
That mailbox story is exactly the gap. The token reached more than the user approved, and nothing in between said no.
On revoking one tool mid-session, I didn't test that live, so I'll stick to what the docs say. Nango sessions can't be changed once created and last up to 15 days by default, so you'd end the session and create a new one. On Scalekit you mint a short-lived session token per run against a Virtual MCP server, so removing a tool from the server takes effect on the next token without touching the connected account. I haven't confirmed whether it also cuts off a token that's already live. For Composio, I'd check with them directly.
I write about Next.js + TypeScript + AI engineering, with a focus on making sites discoverable to LLMs and AI search. Long-form notes at mudassirkhan.me
the questionnaire list in your intro is the real spec. nobody asks about connector count in a security review, they ask who did what, for which user, under which permission.
the one that trips teams up most is the audit trail. per call logs that keep the tool name but not the acting user id are useless the day a customer asks for evidence. we only fixed ours after that exact request landed.
how did each platform behave when a token refresh failed halfway through an agent run? that edge case bit us before any of the headline stuff.
100%. A log with the tool but no user ID won't pass that audit.
On refresh failing mid-run, I didn't test it on all three, so I won't guess for Composio or Nango. On Scalekit, the SDK returns a specific REAUTHENTICATION_NEEDED error when a connection has lapsed, separate from the app rejecting the call. So you can stop the run and send the user a reconnect link instead of retrying blindly. Curious how you handled it on your side.
We're a place where coders share, stay up-to-date and grow their careers.
The shift from "does the connector exist?" to "what is the agent allowed to do?" is the whole game now. A valid token proves identity, but it doesn't stop a prompt-injected email from triggering a destructive action. Scoping permissions at the MCP server level per agent role is a much stronger boundary than just hoping the model behaves.
Exactly. The token tells you who the user is, not what the agent should be doing right now. Scoping per role means a prompt-injected "delete this" has nothing to call, so you're not relying on the model saying no.
exactly. relying on the model to "do the right thing" when it gets prompt-injected is a losing battle. if the token literally doesn't have the delete scope, it doesn't matter how clever the malicious prompt is. it's about moving security from the prompt level down to the infrastructure level.
The migration footnote is the bit I'd underline — refresh tokens tied to the issuing OAuth client means switching vendors later isn't a config change, it's re-authenticating every user. That's the kind of cost that never shows up in the catalog comparison and bites hardest exactly when you're big enough to care.
Yep, that's the one people find out about too late. Moving the config takes an afternoon. Asking every user to reconnect is a support project. Worth thinking about who owns the OAuth apps from day one.
Good framing — the real production gate is authorization and auditability, not connector count: per-tenant token isolation, least-privilege scoping, and an audit trail that ties each action to a user and permission are what the security questionnaire actually probes. The same split shows up in local/desktop agents too: stored credentials still need scoped isolation and an action log, and a self-hostable or VPC-only option becomes a hard requirement once data can't leave the boundary. Practically, that means picking the big managed catalog for fast demos, but leaning toward code-you-own, self-hostable integration when compliance and custom auth (RSA keys, domain-wide delegation) matter.
Agreed, and good point on local and desktop agents. Same problem, smaller blast radius, but the scoping and the action log still matter. Big catalog for the demo, then re-check once compliance shows up. That's pretty much the path most teams take.
the question changing from does a connector exist to what can the agent do once connected is the whole story imo. the one that bit us was your number 2. we gave an agent one token per customer and it could read every mailbox that token reached, which was not what the user actually approved.
per call logs tied to a user and an agent id were what finally made the security review boring.
how do the three handle revoking one tool mid session without killing the whole connected account?
That mailbox story is exactly the gap. The token reached more than the user approved, and nothing in between said no.
On revoking one tool mid-session, I didn't test that live, so I'll stick to what the docs say. Nango sessions can't be changed once created and last up to 15 days by default, so you'd end the session and create a new one. On Scalekit you mint a short-lived session token per run against a Virtual MCP server, so removing a tool from the server takes effect on the next token without touching the connected account. I haven't confirmed whether it also cuts off a token that's already live. For Composio, I'd check with them directly.
the questionnaire list in your intro is the real spec. nobody asks about connector count in a security review, they ask who did what, for which user, under which permission.
the one that trips teams up most is the audit trail. per call logs that keep the tool name but not the acting user id are useless the day a customer asks for evidence. we only fixed ours after that exact request landed.
how did each platform behave when a token refresh failed halfway through an agent run? that edge case bit us before any of the headline stuff.
100%. A log with the tool but no user ID won't pass that audit.
On refresh failing mid-run, I didn't test it on all three, so I won't guess for Composio or Nango. On Scalekit, the SDK returns a specific REAUTHENTICATION_NEEDED error when a connection has lapsed, separate from the app rejecting the call. So you can stop the run and send the user a reconnect link instead of retrying blindly. Curious how you handled it on your side.