Commercial Agent Tools: October 31, 2026

1. Governor Gives Each Agent Its Own API Key With Spend Caps, Tool-Call Rules and a Signed Audit Log in Front of Anthropic, OpenAI and xAI

Governor. Governor, from BSVKey and posted to Show HN on October 31, is a self-hosted Node.js proxy that sits between an agent and the model provider; developers change the SDK base URL and swap the provider key for a per-agent key created with daily and monthly caps and an approval threshold. Before each call it reserves estimated cost against the caps so concurrent calls cannot jointly overshoot, then replaces the reservation with the provider’s real usage figures on return. It inspects tool calls inside model replies before the agent sees them, blocking those a rule denies and holding those that need a person, who can approve, approve a smaller amount, defer, or deny, with a hash of what they were shown recorded in a tamper-evident signed log. Other controls include runaway-loop detection, a global kill switch, per-key model allow-lists and fallback models, rate limits, a “trust ramp” that doubles a new agent’s budget each clean day, and Slack or Discord alerts at 50, 80, and 100 percent of caps. It works with the Anthropic and OpenAI SDKs and with Grok models through the OpenAI SDK, and requires Node 18 or newer with no dependencies. Pricing: not published; access is by early-access sign-up. Source Show HN

2. SpecDeck Hands Coding Agents Reviewed Product Specs Over MCP and Takes Their Decisions Back as Proposed Spec Changes

SpecDeck. SpecDeck, posted to Show HN on October 30, is a browser-based spec editor where product and design teams write feature specs and foundational documents (glossary, decision log, technical context) that agents such as Claude Code, Codex, OpenCode, and OpenClaw read through MCP. A published spec acts like a main branch, drafts are branches, and agent-proposed changes are pull requests that a person reviews and publishes; acceptance criteria live in the spec, and agents mark each one met or unmet when they submit. Work is organized as goals that a person or an agent can claim in agent-led, supervised, or human-led modes, with results returned as reviewable diffs. The paid tier adds an assistant for drafting requirements and spotting conflicting decisions with usage included, a shared spec library and templates, and semantic search. The founder describes the product as early. Pricing: free for up to three people per workspace (agents do not take seats); Organizations at $19 per seat per month with a 14-day trial. Source Show HN

3. Palomino Is a Free GPU-Rendered Rust Desktop Workspace for macOS That Runs Claude Code, Codex and Pi Side by Side

Palomino. Palomino, from Appstudio26 and posted to Show HN on October 30, calls itself an “agent development environment” and is built in Rust on GPUI, the framework behind Zed, with GPU-rendered tiles for terminals, browsers, simulators, files, and agent chats arranged side by side in tabs and workspaces. The company states it does not send code, files, prompts, or chats to its own servers; only the AI services the user selects receive content. In the Show HN thread the team said multi-agent orchestration, cloud machines through providers such as Daytona, and remote machine management are coming, and that a more modular UI is planned in response to commenters who objected to a closed-source harness. Updates ship through an in-app check. It is Mac only, with Windows and Linux on a waitlist. Pricing: free, with unlimited workspaces, agents, and panes and no account required; the business model is not stated. Source Show HN

4. Kitepod Is a Free Hosted WordPress MCP With 51 Tools That Stages Risky Edits and Routes Code and Database Changes to Human Approval

Kitepod. Kitepod, posted to Show HN on October 30, connects Claude, ChatGPT, Cursor, or Claude Code to any number of WordPress sites by installing a plugin (WordPress 6.4 and PHP 7.4 or newer) and adding its hosted endpoint as a connector; the assistant acts as a chosen WordPress user and never gets more power than that user. Its 51 tools cover pages, posts, media, menus, global styles, templates, and theme files edited in a previewable draft, native page-builder formats (Elementor, Divi, Bricks, Beaver Builder, and others), WooCommerce products and coupons, plugin and theme installs with a checkpoint before each, and WP-CLI and database access where reads are free but writes, search-and-replace, and PHP need approval. Every write is snapshotted for 90 days with undo per edit, task, or session, and after a change goes live Kitepod screenshots the page before and after. Agencies can invite clients as approvers for their own sites. Pricing: free, with unlimited tool calls and sites and no card required; the page does not explain the business model. Source Show HN

5. Opencommerce Gives Agents One Buy Tool Across More Than 5,000 Online Stores for 1% of Each Completed Order

Opencommerce. Opencommerce, from Agentcard and posted to Show HN on October 30, positions itself as “OpenRouter for commerce”: a single endpoint exposed as a buyTool in the @agent-cards/commerce-sdk npm package that plugs into Vercel AI SDK-style tool runners alongside Claude, GPT, or Gemini models. The flow is search, add to cart, confirm cart, request payment approval through a link sent to the shopper, after which Agentcard places the order and streams updates; no merchant integrations or agent-driven browser are needed. The company claims coverage of 5,026 stores, search-to-purchase in under 20 seconds, and a warmed route of about 9 seconds versus 2 minutes 40 seconds cold, and it runs a public monitoring page that exercises the search-to-checkout flow against each merchant and publishes 45-day success histories. A wallet integration is listed as coming soon. It is a different company from StoreRouter, covered on October 28. Pricing: 1% of each completed order on the standard plan with no API access; an enterprise plan adds API access at custom fees. Source Show HN

6. StrictLLM Is a Desktop Chat Client That Enforces Per-Persona Tool Permissions in Code and Sandboxes Local MCP Servers in Podman

StrictLLM. StrictLLM, posted to Show HN on October 29 after nine months of development, is a Mac, Windows, and Linux desktop app whose central claim is that every tool or MCP call is gated by code rather than prompt instructions: each persona gets Always Allow, Always Ask, or Always Deny per tool, and local MCP servers run isolated in Podman containers with configurable network access. Multiple models can share one conversation thread, with personas pulled in by @-mention, against local or cloud backends over OpenAI- or Anthropic-compatible APIs including LM Studio, Ollama, Lemonade, OpenRouter, Claude, and ChatGPT. The vendor states it collects no telemetry and that the only network calls are update checks and license activation. The author contrasts it with “manual mode” in other tools, which they say does not guarantee a prompt before each tool call. Pricing: $29 per month for three personas without orchestration and a seven-day offline grace period, or $79 per month for unlimited personas, full orchestration, and full offline mode, plus enterprise contact; a trial key for Show HN readers runs through November 30. Source Show HN

7. Beolis Runs Coding Agents in Parallel on Local Repos, Worktrees or Remote Workers Through Scripted Workflows

Beolis. Beolis, posted to Show HN on October 30, manages coding agents running in parallel or in the background on the local machine (the current repository or git worktrees) or remotely, where it creates a worktree and runs the agent there. Its distinguishing idea is that users or agents write code for the default workflow steps rather than prompting at each step: a script runs the tests, a cheap agent summarizes failures, and a stronger agent is invoked only to fix them, which the author says stretches token budgets. The author has built it since February, used it to build itself, and describes it as stable for daily use. The landing and pricing pages require JavaScript, so detail beyond the author’s Show HN description could not be verified. Pricing: free for local use, with paid tiers for collaboration features and remote workers at unpublished prices. Source Show HN