Reference · Landscape

Aggrete vs the other MCP security tools.

MCP security tooling falls into a few honest categories, and they mostly solve different problems. This is where each fits, and where Aggrete is genuinely different, without pretending it does everything.

Scanners (mcp-scan, truecopy, and similar) inspect a server or tool before you install it and flag known-bad or poisoned definitions. Essential hygiene, but they run once, at install, and do not govern what happens at runtime.

Model guardrails classify prompts and outputs at the edge. They catch a lot, but they are probabilistic: a novel injection is a coin flip, and they add a model to your decision path.

Gateways and proxies (docker/mcp-gateway, agentgateway, IBM mcp-context-forge) route traffic, authenticate, and often rate-limit. Excellent plumbing. Their policy is usually a per-call allow or deny: this tool, this user, yes or no.

Aggrete is a proxy too, so it sits in the same spot as a gateway, but the policy is different in kind: deterministic, written as your code of conduct, and stateful. It reasons over the whole session, so it refuses the forbidden combination and the toxic flow, not just the single call. And it never puts a model in the decision.

Static scannerModel guardrailGateway / authzAggrete
Runs atinstall timethe edge (prompt/output)each callevery call, in front of the upstream
Decides witha catalog of known-bada model classifierrouting and authz rulesa deterministic policy rule
Stateful across callsnononoyes
Refuses forbidden combinationsnopartialnoyes
Stops the lethal trifectapartialpartialnoyes
Tool integrity (poisoning, rug pull)yesnopartialyes
Per-user identity to the upstreamnonopartialyes
Deterministic (same answer every time)yesnopartialyes
Policy owned by non-engineersnononoyes

What Aggrete does not do

Being honest is the point of a comparison. Aggrete is not a content classifier: if you want a model to judge toxicity or PII in free text, pair it with a guardrail. It is not a package-reputation database; a scanner's catalog of known-bad servers is complementary, and Aggrete happily runs behind or in front of one. And it is a policy layer, not a full API gateway: if you need advanced traffic shaping, run it alongside your gateway. The two compose, they do not compete.

Where Aggrete wins is the problem the others structurally cannot touch: the risk that only exists across several calls, decided by a rule you can read, owned by the person who wrote the policy, enforced before anything is fetched.

Judge for yourself. The runnable attack repros are in examples/attacks, and uvx aggrete --demo gives you a governed sandbox in one line.
Open source

Govern what your assistants can reach.

Aggrete is Apache-2.0 and runs on a laptop or a cluster. Star it, fork it, and tell us what rule to build next.

Star on GitHub Beginner's guide How it works