What it does
Aggrete sits between MCP clients and upstream connectors. It checks each call against a YAML policy and keeps per-user memory across calls and sessions, so a request that is allowed on its own can still be refused when it completes a forbidden set of data. Results are redacted for emails, SSNs, card numbers and tokens before they reach the model, and tools a user cannot call are hidden from that user. It also refuses writes made after a session has read untrusted content, flags hidden instructions in tool descriptions, holds upstream credentials itself, and writes a hash-chained audit log of every decision.
It suits teams that connect several MCP servers to one assistant and need rules that look at combinations of access rather than single calls. A scan mode inspects an existing MCP setup for cross-server leak paths, poisoned descriptions, secrets in client configs and servers that bypass the proxy, and can write a suggested config and policy. Actions can be held until a named clause owner approves them, from Slack, the terminal or the console. A conformance command runs sixteen checks mapped to the OWASP MCP and Agentic Top 10 lists, CoSAI and AIUC-1.
Setting it up involves writing the policy in coc.yaml, optionally drafting it from a handbook with an ingest tool. Remote use runs over streamable HTTP with OAuth, and per-user credentials can be configured. An adapter mode runs the engine inside an existing gateway such as agentgateway, Docker MCP Gateway or IBM ContextForge. A Claude Code agent skill can guide setup and operation.
Limitations and requirements
Denial after a call redacts results but does not prevent the fetch, so rules should be decidable before the call. Identity over stdio is advisory; real enforcement needs streamable HTTP with OAuth and blocking of direct connector grants at the identity provider. Aggregation limits can be narrowed but not fully closed: a user who spaces requests beyond the window or paraphrases across systems the proxy does not front can get through. The project is not a gateway and does not provide multi-tenancy, a token vault or high availability, so production use is expected to run behind a gateway.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.Aggrete/aggrete
MCP policy proxy: enforces a code of conduct across connectors before the upstream is called.
v0.8.3As shown on registry.modelcontextprotocol.io when we last read it - Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.aggrete/aggrete
MCP policy proxy: enforces a code of conduct across connectors before the upstream is called.
v0.13.4As shown on registry.modelcontextprotocol.io when we last read it