Rust single-use claim gate stops tool replay and duplicate execution in local LLM workflows
AdmissibilityScience · reddit · 2026-09-22
Running local models via Ollama and LM Studio with custom system scripts, the author found the real pain wasn't the initial tool call but duplicate retries and mid-sequence parameter mutations when a model gets stuck in a loop.
He built a lightweight Rust CLI/library gate (v0.1) where the local model acts purely as an untrusted proposer:
- A trusted issuer creates a single-use claim bound to action + resource + payload SHA-256
- The agent submits its tool request
- The first matching request executes and atomically burns the claim on disk (OEXCL / createnew)
- Any replay or mutated payload immediately gets a DENY
Currently a local CLI tool and library function (executerequest()), not a full stdio MCP proxy—you call it inside tool handlers before firing file modifications or shell scripts. Open source: arcstone-mcp-sidecar on GitHub.
Related event: Developer Releases Rust Sidecar to Block MCP Tool Replay Attacks(2 posts)→
More from coding & agent
- Agent Memory Challenge 2026 Cycle 2 opens with $22K prize pool and 150-task coding track — omarsar0 · 2026-09-22
- dotey shares doc-driven multi-agent coding workflow: Fable writes spec, Opus executes — dotey · 2026-09-22
- One LLM call decides what to remember: building fast, cheap agent memory on Jev — gdechichi · 2026-09-22
- 8 open-source agent tools: computer-use, browser control, memory and more — Aiden_Tech_Ai · 2026-09-22
- ChatGPT allegedly emailed the FBI on its own, joining a week of rogue agent incidents — JanJanJaJa · 2026-09-22
- Should SaaS be built for agents instead of humans? A dev ships uivoid to test it — Corridl · 2026-09-22