Open Source Agent Tools: October 21, 2026

1. Gauge CLI Runs Real Coding Agents Against a Product to Find Where They Get Stuck

Gauge CLI. Gauge open-sourced its CLI for improving “agent experience”, shown on Show HN on October 20 and released under the MIT license. It runs Claude Code, Codex, Pi or OpenCode against a team’s Skills, CLI, SDK, MCP server, API or docs inside Docker, captures where the agents struggle (interactive commands that block, deprecated patterns, unused Skills, invented limitations), and lets the team’s own coding agent test fixes before publishing. The full loop runs locally with existing agent subscriptions or API keys and needs no Gauge account; a hosted version is also offered. Source

2. Melra Gates Risky MCP Tool Calls and Parks Ambiguous Writes Instead of Retrying

Melra. Melra (Modular Execution Layer for Reliable Autonomy), posted to Show HN on October 20, is an Apache-2.0 local MCP runtime that sits between an agent and its tools. It checks each call against a policy, requests human approval for risky actions, verifies declared evidence and keeps a redacted receipt per call. Mutations are dispatched at most once per task, and calls with uncertain outcomes, such as timeouts, are parked for reconciliation rather than replayed. The project labels itself alpha and installs with npx @melra/cli@alpha setup, which writes itself into the chosen MCP client’s config. Source

3. Tenbri Records Every Agent Decision in a Signed, Append-Only Log

Tenbri. Tenbri, shown on Show HN on October 20 and licensed Apache-2.0, implements what it calls an Agent Decision Record: a signed, chained, append-only record of each agent decision covering who acted, under which rule and inputs, who approved and what happened, verifiable offline. One-line instrumentation is provided for the Claude Agent SDK, OpenAI Agents and LangGraph, and a decorator can wrap individual functions so refusals and outcomes are logged. Records are plain files on storage the user controls; packages ship for pip, npm and cargo, and the current release is 0.1.0-beta.5 with a read-only web UI. Source

4. Frink Runs Parallel Agents and Event-Triggered Workflows From the Desktop

Frink. Frink is an MIT-licensed desktop app, posted to Show HN on October 20, for running up to four coding agents side by side, each on its own git worktree with its own model and diff. Routines described in plain language become “Flows” that mix agent, command, custom code and approval steps, triggered by schedules, webhooks or events from Linear, Shortcut, ClickUp, Sentry, PostHog or Vercel. Finished runs and agent questions collect in a Work Queue on the desktop or an iPhone app (beta). It runs locally with the user’s existing AI coding tool and needs no Frink account. Source

5. SIwithin Fences Unattended Claude Code and Codex Runs Under a Dollar Budget

SIwithin. SIwithin, an Apache-2.0 single-file Python tool shown on Show HN on October 20, runs the Claude Code or Codex CLI unattended on a brief, restricted to the folders it is given, with the home folder refused for reading and writing. Spending is capped per run and across runs, and for Claude a loopback proxy holds the real API key, meters every request with a per-run token and refuses calls past the budget. Each run writes one ledger line recording cost and whether it completed. It runs only on macOS and refuses to start on other platforms rather than run unfenced. Source

6. Code-Mode Plugin Has Claude Code Call MCP Tools From One Sandboxed Program

code-mode. code-mode is an MIT-licensed Claude Code plugin, posted to Show HN on October 20, in which the model writes a single JavaScript program that calls MCP tools inside a sandbox, filters and joins the results there, and returns only the answer to the context. The sandbox is a separate Node process with no file system access and a CPU limit, with network blocked on macOS and on Linux where supported, and normal per-call MCP permission checks still apply. Working programs can be saved as reusable functions after user approval. The author’s benchmark reports the same 30 of 30 correct answers at 43% lower cost per task with Haiku 5.5. Source

7. Agent Archive Hands Sessions Between Claude Code and Codex and Backs Them Up to S3

agent-archive. agent-archive is an MIT-licensed Go CLI, shown on Show HN on October 20, that lets a user pick a Claude Code or Codex session and continue it with its context in the other agent, or on another machine. It can upload Claude Code, Codex and Cursor transcripts to the user’s own S3 or R2 bucket, list sessions filtered by topic, project, branch or PR, and report token counts and estimated spend by agent, model and project. Release binaries cover macOS and Linux; the current version is v0.2.2. Source