Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Agent tooling splits in two: runtimes that execute, scanners that audit

Five open-source agent projects posted to Show HN show the same split: runtimes that execute agent work, and security tooling that watches it.

AI & modelsExplainerGrace OkonkwoPublished: 28 September 20264 min readSources 5
Agent tooling splits in two: runtimes that execute, scanners that audit

Agent tooling is settling into two jobs. One runs agents long enough to be useful. The other finds out what they can reach on your machine. The five projects below, posted to Hacker News' Show HN, sit at different points on that line.

Runtimes

Pizza Bot is the clearest example of the first job. Its GitHub README describes a local-first inbox for long-running AI agents, built on DeepAgents and LangGraph. Start or schedule a task, walk away, and return to finished output in an Unread queue. Approval requests land in a separate Action queue. Agents keep working when you leave the page or disconnect. Only the api-server process has to stay running.

The project was developed at Amazon and is released under the Apache 2.0 license, according to its repository. It supports Amazon Bedrock, Anthropic, Google Gemini, OpenAI, OpenRouter and Ollama. Two details matter for anyone evaluating it. First, checkpointed runs survive client disconnects. Cron or webhook triggers can start work without an open conversation. Second, local file access is opt-in: folders are added individually under Settings. The README states Pizza Bot receives no default home-directory access.

Soma, documented at docs.trysoma.ai, takes a different angle. It is a self-hostable agent and workflow runtime that ships as a single binary. It is aimed at startups through to enterprise, according to its documentation. Code is written in TypeScript, with Python listed as coming soon. The runtime handles the plumbing around it. Execution is fault-tolerant and resumable: it can crash or suspend and pick up where it left off. It also generates A2A and OpenAI Streaming compatible endpoints automatically.

Its platform table is unusually candid. TypeScript and Python builds are marked available on macOS Intel and ARM and Linux x86 and ARM. Windows is listed as planned but not natively supported, because the project uses Unix domain sockets in Rust. Rust itself is marked unavailable across every platform in the table.

Registries and the developer loop

Artifact Keeper is not an agent runtime at all. It lands in the same conversation because agents need somewhere to pull packages from. The GitHub organisation page describes a self-hosted artifact registry positioned as a drop-in replacement for JFrog Artifactory and Sonatype Nexus. It has no open-core split and no enterprise edition. It claims native protocol support for more than 45 package formats, so pip, npm, docker, cargo, helm and go talk to it directly rather than through a generic blob store.

The project ships a backend in Rust with Axum, SQLx and Wasmtime, a Next.js dashboard, native iOS and Android apps, and a Helm chart with Terraform modules. Security features include Trivy scanning, SBOM analysis through OWASP Dependency-Track, GPG and PGP signing, and SSO via OpenID Connect, LDAP, SAML 2.0 and JWT. A built-in migration tool moves repositories, artifacts, users and permissions from JFrog Artifactory. The quickstart is three commands: clone, cd, docker compose up.

Claumon is narrower but more pointed. Its README describes a local dashboard for Claude Code that forecasts usage limits from a published empirical-Bayes model. The model is refit daily on the user's own history, with an 80 percent credible interval and an ETA to threshold. Rate-limit gauges come from the Claude OAuth usage API. The README contrasts this with trackers that estimate limits from local logs.

The project's argument for existing is that Anthropic's usage analytics dashboard is aimed at Team and Enterprise org admins, not individual Pro or Max subscribers. Personal /usage pages show current standing with no history. Claumon also manages the ~/.claude directory: per-session costs, SQLite-backed daily aggregates over a 7 to 90 day window, running-process control, and a memory-file browser. It binds to IPv4 loopback only. The README recommends an authenticated SSH tunnel rather than exposing port 3131.

The audit layer

Golf Scanner covers the second job. It is a single static Go binary, with three dependencies, that discovers MCP server configurations across seven IDEs: Claude Code, Cursor, VS Code, Windsurf, Gemini CLI, Kiro and Antigravity. It runs 20 security checks, nine offline and 11 online. These query OSV, GitHub, npm, PyPI, OCI registries and the MCP Registry, then produce a 0 to 100 risk score per server.

The scoring rules are published. The worst severity per check sets a 0 to 10 score. Scores combine as a severity-weighted average (critical 10x, high 7.5x, medium 5x, note 1x). Hard caps apply: any critical finding caps the total at 30, any high finding at 59. The README claims zero telemetry and no account requirement. Online checks cover command safety, plaintext credentials, world-writable scripts and binaries, container isolation, typosquatting, digest pinning and cosign signature verification.

Comments 0

Sources

5
  1. 01Show HN: Pizza Bot – An inbox for AI agents that work in the backgroundEN
  2. 02Show HN: I built an open-source Rust/TS AI agent runtime with a Next.js-style DXEN
  3. 03Show HN: Artifact Keeper – Open-Source Artifactory/Nexus Alternative in RustEN
  4. 04Show HN: Golf Scanner – OSS tool to find and audit every MCP serverEN
  5. 05Show HN: Claumon – forecasting Claude Code usage limits with a Gamma processEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.