Open Source Agent Tools: October 31, 2026

1. Fastbrowse Picks Browser Actions From a Page’s Indexed Controls With a Small Choice Model Instead of Screenshots

Fastbrowse. Fastbrowse, from Agent Labs and posted to Show HN on October 30, is a Python CLI and library with a Rust core that drives Chrome by indexing the controls present on each page and having a small “choice” model (Jev, from Typesafe) pick one, so the agent cannot click elements that are not there and there are no selectors to maintain. A larger LLM, reached through OpenRouter, Vercel AI Gateway, or a fallback such as Gemini Flash Lite, handles planning, reading, and writing; finished answers cite verbatim page quotes with scroll links, and callers can pass a schema for validated output. Secrets are passed by name and typed only on the declared origin, irreversible actions such as submit, pay, delete, and send pause unless --authorize is set, and step, time, and dollar limits are enforced in code. It runs against local Chrome, any running Chromium or Electron app, or a Browser Use Cloud browser, and ships as a CLI, Python library, Node and Go SDKs, an MCP server, and a coding-agent skill. The author’s 46-task, three-run comparison against Browser Use’s hosted agent reports a 99.2% pass rate versus 96.3% at $0.0296 versus $0.318 per run, though slower per task (18.9 versus 15.3 seconds); the numbers are self-reported. Requires uv and Python 3.13 or newer. License: MIT. Source Show HN

2. Imza Is a Native C++ Terminal Coding Agent Whose Only Tool Is a Sandboxed Lua Interpreter

Imza. Imza, posted to Show HN on October 31 with v0.7.1 shipped on October 27, is a coding agent written entirely in C++ with a native terminal UI, a roughly 20 MB binary, and 5 to 8 MB of RAM at startup. Work moves through three modes: Plan keeps the repository read-only and produces a co-editable Markdown spec with pinned notes and kept revisions, Build implements the plan while streaming every change as a diff, and Review lands the whole change in a code-review view whose findings feed a new plan revision. The model receives a single tool that executes Lua scripts, with C++ methods exposed as Lua bindings behind a permission layer, so several operations can be orchestrated in one call and all filesystem access is mediated. It supports MCP, skills and instruction files, more than 50 providers including OpenAI subscriptions and local OpenAI-compatible models, up to eight concurrent subagents, /undo, a side chat, headless CI execution, and overridable prompt fragments; no account is required. License: GPL-3.0. Source Show HN

3. mobile-agent Drives Android Devices, iOS Simulators and iPhones for Agents From a Rust Daemon That Mirrors Each Screen

mobile-agent. mobile-agent, posted to Show HN on October 30, is a Rust daemon with a CLI and an MCP server built around the latency of each tap-and-observe step: it keeps a live mirror of every attached screen (windows, accessibility tree, latest frame), updates it from platform change events, and answers snapshots, waits, and settle checks from that mirror instead of polling and sleeping. On Android a single-dex agent is pushed over adb and runs under app_process as the shell user holding the UI-automation role, with no APK or instrumentation. On the iOS Simulator, input goes through dtuhidd, frames come from IOSurface, and the accessibility tree comes from a Rust reader spawned into the Simulator with no XCTest runner; physical iPhones need a Rust XCTest runner built from the repo and a CoreDevice tunnel, without root. The CLI follows agent-browser’s command grammar, element references, and JSON envelope so existing prompts and skills carry over. The author calls it experimental and pre-release: it must be built from source, relies on private Apple and hidden Android interfaces, needs a macOS host for iOS (Linux hosts get Android only), and requires Xcode 26 or 27. License: MIT. Source Show HN

4. Offense Gives Coding Agents Security Scanning Over CLI, MCP and HTTP With a Free Local Dependency Engine

Offense. Offense, from Offense Research and posted to Show HN on October 30, returns structured security evidence (a scan.json plus a remediation.md) for a coding agent such as Claude Code, Codex, Pi, Hermes, or Crush to act on with its own model and repository permissions. It runs in three modes: a local dependency engine that queries public OSV and npm metadata with no account, a self-hosted platform for Linux amd64 under Docker with a Node API, console, SQLite, and isolated Docker scanners (dependency audits, Opengrep 1.32.0, and Zizmor 1.31.1 enabled, billing disabled), and a hosted service where paid tools draw on quoted prepaid credit. A one-click PR installer adds a pinned, read-only GitHub Actions dependency audit, an agent SKILL.md, and recipes for PR review, secrets checks, SQL-injection review, and Actions YAML review. The README notes that local mode still needs network access for advisories, the self-hosted profile binds to loopback and is not multi-tenant, and other catalog tools stay disabled outside the hosted service. Requires Node 24 or newer. License: Apache-2.0. Source Show HN

5. Bugbout Benchmarks Coding Agents on Bugs Your Own Repository Already Fixed, Scored by the Maintainers’ Hidden Tests

Bugbout. Bugbout, posted to Show HN on October 29 and installable with pip install bugbout, mines a project’s git history for fixes that reference an issue and add a test, establishes ground truth by running the maintainers’ tests before and after each fix, then gives each agent the issue and a history-free checkout of the pre-fix code inside an OS sandbox with web access off. Codex, Claude Code, Gemini CLI, and mini-swe-agent work out of the box, and any CLI agent can be added through bugbout.toml. Every test an agent writes is run against every other agent’s patch, a policy picks one patch per bug, and the hidden tests score each patch and each pick. Bouts are resumable, adding an agent later runs only that agent, and a Python API and Markdown reports are included. Python 3.12 or newer on macOS or Linux (WSL2 on Windows). License: MIT. Source Show HN

6. dbprint Snapshots a Database’s Schema, Joins and Column Statistics Into Git-Committed Text That Agents Read Offline Over MCP

dbprint. dbprint, posted to Show HN on October 30 and runnable with uvx dbprint, measures a database once (DDL, declared and inferred joins, per-column null shares, full value lists with every spelling, ranges, and value shapes) for PostgreSQL, MySQL, Snowflake, BigQuery, Databricks, Redshift, ClickHouse, or DuckDB and writes the result as YAML and Markdown next to the code. An annotate command hands an agent one table’s measurements and open questions so it can document meaning (sentinel dates, dead columns, gross versus net) in files dbprint never overwrites, and a --check flag flags notes the data contradicts. Later sessions read the print through an MCP server, a packaged skill, or dbprint context, so no database connection or credentials sit with the agent, and dbprint check gates schema or data drift in CI. Values can be masked, excluded, or hashed by fully qualified name or inferred sensitivity. The author’s internal tests claim 80% fewer wrong answers and 17% lower cost per correct answer; these are self-reported. License: Apache-2.0. Source Show HN

7. Navital Gym Grades Agents Deterministically on Indian-Context Tasks Such as Rail Trip Planning, UPI and GST Invoicing

Navital Gym. Navital Gym, from Navital Labs and posted to Show HN on October 30, ships six sandboxed task sets: trip planning under train quotas and waitlists, UPI payments, GST invoicing and returns, single and multi-document form filling, and e-commerce support in Hindi and Hinglish, with Tamil, Bengali, and Marathi as labels. Every task is graded by deterministic rules that recompute results from the sandbox world rather than by an LLM judge; nothing the agent claims is trusted, many tasks are deliberately infeasible (17 of 55 trip tasks) and the agent must say so with the right reason, and a failing test enforces that expected answers never reach the agent. The core and environments run on standard-library Python 3.10 or newer, with optional adapters for Inspect AI, verifiers, and OpenEnv; tasks are also on Hugging Face and a public leaderboard is maintained. The README stresses that the rules are simplified and synthetic and that the project is unaffiliated with NPCI, RBI, GSTN, or Indian Railways. License: Apache-2.0 for code, with a separate data license. Source Show HN

8. Archwhere Puts Every Coding-Agent PR on a Declared Architecture Map That the Agents Must Keep Current to Merge

Archwhere. Archwhere, posted to Show HN on October 31 hours after its repository was created, is a zero-dependency JavaScript CLI (npx archwhere) whose model is a single JSON file mapping globs to components with a kind (security, product, ops) and edges between them. init drafts the model from directory structure, workspaces, and import edges, then prints a prompt for the coding agent to refine it; check fails CI when files are unmapped or globs are stale, and check --fix adopts files into the nearest component. A GitHub Action comments on every PR with a component, kind, and files table, security-zone flags, and a Mermaid system map with touched components highlighted. A mission command serves a local control UI on port 4610 with an attention rail (security, blocked, needs attention, ready to merge) and threads pinned to components that agents answer with PRs, and a SKILL.md tells agents how to maintain the map. It is brand new from a solo developer and demonstrated on a fictional product. License: MIT. Source Show HN