xagi-lab/melra is an open-source local MCP runtime that inspects AI agent tool calls against policy and routes risky actions through approval. MCP (Model Context Protocol) is the protocol that defines how agents connect to external systems. The repository was created on July 28, 2026 and has no public release yet. According to the README, every call is checked against policy, risky actions run only after an approval step, and execution results are recorded as receipts with the values redacted.
The core of this runtime is not retrying calls whose outcome is ambiguous. A write operation whose execution status cannot be known after a timeout is not resent as it is; it is held as a pending reconciliation item so that the result can be verified later. Mutating actions are stated to be delivered exactly once per operation. The rationale declared by the agent is reviewed, and whether the execution result matches is checked as well. The implementation language is TypeScript and the license is Apache-2.0.
The boundary is clear. The README states that whether the effect happens exactly once in the external system is left to the provider. Not retrying and having the target system process a call exactly once are two different problems. The repository has 11 stars, and in the evaluation score only momentum remains, at 0.2 out of 10, while freshness, technology and quality are empty. Rather than deciding on adoption now, it is closer to a project to keep an eye on.