Releases·Tool Release
Show HN: Agent Chaperone – Screen AI agent tool calls and results with Jev
A developer released agent-chaperone, an open-source screening layer for AI agent tool calls and tool results, inspired by TypeSafe's Jev release the prior week. It blocks risky, off-task, or secret-leaking actions before execution and screens incoming tool results for prompt injections before the agent reads them. It gives practitioners a guardrail layer for running agents against untrusted tools and data.
3 engagement1 sourceMon, Sep 21, 2026, 04:34 PM
AI Overview
- agent-chaperone is a screening layer built on TypeSafe's Jev that blocks risky, off-task, or secret-leaking tool calls before they run.
- It screens prompt injections and suspicious instructions in tool results before the agent reads them; code is at github.com/agent-chaperone.
Entities
Related
Releases·Tool Release
Show HN: Deterministic Guardrail That Allows or Blocks AI Agent Commands, Now on GitHub
A developer launched a deterministic guardrail tool for AI agents on GitHub via Show HN, allowing or blocking agent commands (e.g. permitting 'terraform plan' but blocking 'terraform destroy') to prevent destructive actions. The author explicitly labels it a first proof of concept, not a hardened sandbox, and is soliciting community feedback on usefulness and enterprise interest.
A developer launched a deterministic guardrail tool for AI agents on GitHub via Show HN, allowing or blocking agent commands (e.g. permitting 'terraform plan' but blocking 'terraform destroy') to prevent destructive actions. The author explicitly labels it a first proof of concept, not a hardened sandbox, and is soliciting community feedback on usefulness and enterprise interest.
3 engagement1 sourcehackernewsFri, Oct 16, 2026, 11:53 AM
Products·building
Local runtime gate that intercepts and vets AI agent tool calls before execution
An open-source pilot that sits between an LLM agent and its tools, intercepting each tool call at runtime. It checks intent, simple data provenance, and policy, then returns ALLOW or BLOCK with a reason and writes an audit log. It targets developers protecting agents from prompt-injection-driven tool calls (e.g., a hidden instruction in a document triggering unintended HTTP, email, DB, or file actions).
An open-source pilot that sits between an LLM agent and its tools, intercepting each tool call at runtime. It checks intent, simple data provenance, and policy, then returns ALLOW or BLOCK with a reason and writes an audit log. It targets developers protecting agents from prompt-injection-driven tool calls (e.g., a hidden instruction in a document triggering unintended HTTP, email, DB, or file actions).
10 engagement1 sourceredditWed, Sep 23, 2026, 01:00 PM
Social·Community
TypeSafe's Jev bets on language-model-free AI for small decisions
The Emerging Stack profiles TypeSafe's Jev, which bets there is a better approach than using highly capable language models for tiny decisions. The piece argues current AI usage over-applies conversational LLMs to tasks that may not need them.
The Emerging Stack profiles TypeSafe's Jev, which bets there is a better approach than using highly capable language models for tiny decisions. The piece argues current AI usage over-applies conversational LLMs to tasks that may not need them.
12 engagement2 sourcesrss, redditFri, Sep 18, 2026, 04:57 AM
Social·Community
Reddit pilot: Gemma 4 E2B's verbalized confidence sits on 0.00/1.00/0.50 for 99.6% of calls; Jev judge ranks passages better, AUROC 0.926 vs 0.804
A developer ran a 500-question SimpleQA pipeline where Gemma 4 E2B writes answers and a judge makes retrieval and verification decisions, comparing Gemma's verbalized self-confidence against TypeSafe's Jev API judge. Gemma's confidence collapsed onto three values (99.6% of judgments at 0.00, 0.50, or 1.00) and ranked passages worse than Jev (AUROC 0.804 vs 0.926), yet the fully judged pipeline still underperformed a no-judge baseline (0.612 vs 0.740) because it declined too many answerable questions.
A developer ran a 500-question SimpleQA pipeline where Gemma 4 E2B writes answers and a judge makes retrieval and verification decisions, comparing Gemma's verbalized self-confidence against TypeSafe's Jev API judge. Gemma's confidence collapsed onto three values (99.6% of judgments at 0.00, 0.50, or 1.00) and ranked passages worse than Jev (AUROC 0.804 vs 0.926), yet the fully judged pipeline still underperformed a no-judge baseline (0.612 vs 0.740) because it declined too many answerable questions.
13 engagement1 sourceredditMon, Sep 21, 2026, 06:16 PM
Products·launched
Unofficial Jev plugin for coding agents: best practices, API reference, pattern library, and evals
A plugin that lets LLM coding agents (any harness) work with Jev more effectively by providing best practices, anti-patterns, an API reference, and a catalog of 150+ community projects grouped by design pattern with code sketches. It uses progressive disclosure so agents load only the context needed for the current task. The author also ran repeated evals comparing the plugin against no plugin and TypeSafe's official plugin.
A plugin that lets LLM coding agents (any harness) work with Jev more effectively by providing best practices, anti-patterns, an API reference, and a catalog of 150+ community projects grouped by design pattern with code sketches. It uses progressive disclosure so agents load only the context needed for the current task. The author also ran repeated evals comparing the plugin against no plugin and TypeSafe's official plugin.
4 engagement1 sourceredditSun, Sep 27, 2026, 03:55 AM