New Developer Tools Worth Knowing — September 2026

Our submission form gets a steady trickle of tools. Most get a polite no — astrology apps, photo upscalers, marketing suites, things with no connection to building software. A smaller number are squarely in our lane, and this is the first monthly round-up of the ones that survived a look.

What “survived a look” means here, precisely: I checked the claims I could check, on the date below, against the vendor’s own repo or docs. I have not run any of these through real work. Where something is unverified or unclear, it says so. None of these companies paid anything, all five arrived through the free submission form, and none of them saw this before publication.

HOL Guard — a permission gate for coding agents

The problem it targets is one anybody running agents has thought about at 2am: your agent can run shell commands and install packages inside a single session, and you are trusting a model’s judgement about what is safe. HOL Guard sits in front of those actions and blocks the risky ones — shell commands, file access, package installs, MCP tool calls — checking for secret exposure, destructive operations, prompt injection and malicious dependencies.

It hooks into Codex, Claude Code, GitHub Copilot CLI, Cursor, Cline and Gemini CLI among others, through native agent hooks and MCP proxies, with a support matrix documenting what is actually enforced per agent — a detail I appreciated, because coverage varying by agent is exactly the sort of thing that gets glossed over. Install is pipx install hol-guard && hol-guard init, Python 3.10+.

Checked 20 September 2026: 635 stars, 79 forks, Apache-2.0, actively maintained. Approvals and decisions stay local by default; an optional paid cloud adds team policy. Sources: repo, product page, and the live support matrix — which is worth a look in its own right: per-agent coverage with dated verifications and commit evidence.

Worth knowing: it comes out of Hashgraph Online DAO, a crypto-adjacent organisation. The tool itself is agent security with no chain involvement that I could find, but you should know whose project it is.

Skilled — which agent skills you actually use

A terminal dashboard that reads local session history from Claude Code, OpenCode, Codex, Grok and Droid, and shows you what you actually reach for: bar charts, activity heatmaps, hourly histograms, and an audit that flags trending versus stale skills. If you have accumulated more skills than you can remember, it answers the question directly.

The privacy claim is unusually clean for this category: “No data leaves your machine. No accounts, no config files, no API keys.” Zero network, zero telemetry, local history files only.

Checked 20 September 2026: 48 stars, MIT. Installs via npm (@avcodes/skilled), pip, or a shell script. Source: repo.

Pragor — a shared board for humans and agents

Aimed at teams running agents from different frameworks — CrewAI, LangGraph, AutoGen — that need somewhere to meet. One board carries messages, tasks, files, approvals and a calendar; agents connect over REST or MCP; risky operations sit behind human sign-off, and there’s an audit trail of who did what.

Checked 20 September 2026: the beta free tier is 3 projects, 10 agents, unlimited activity events, 100 MB storage, up to 3 board members, 1 agent wake daily. After beta the standard free tier drops to 1 project, 3 agents and 100 events a month. Paid plans are priced but not yet purchasable — Solo €21.75/mo, Team €74.25/mo, Business €186.75/mo, with event packs from €19. Source: pricing page.

Worth knowing: it’s closed-source SaaS, and the numbers above will move when beta ends. If the beta allowances are what attract you, note which ones are temporary.

FrameThrower — a film-reference API

The odd one out, and the reason it’s here is the shape of it rather than the subject. It’s a searchable library of frames from thousands of films, and the submission was specifically for the developer surface rather than the consumer product: eleven REST endpoints, an npm package (framethrower-ai), and an MCP server that plugs into Claude Code, ChatGPT and Cursor via claude mcp add --transport http framethrower https://framethrower.ai/api/mcp.

Checked 20 September 2026: free signup with no card, $2 of credits for MCP usage, then pay-per-use. REST uses token auth, MCP uses OAuth 2.1. Source: framethrower.ai (the API docs themselves now sit behind a login).

Worth knowing: their marketing names large studios as users. I have not verified those relationships and am not repeating them as fact.

KivTools — browser utilities, a lot of them

A large collection of the small utilities you keep a tab open for: JSON formatting and validation, OpenAPI and GraphQL exploration, JWT inspection, hashes, Base64, DNS and SSL lookups, PDF and image manipulation, code formatters.

Checked 20 September 2026: claims 315 tools across 19 categories. Free, no paywall messaging. The site states “many tools run locally; network checks send requests as explained on each tool page” — which is the honest way to put it, and worth reading per-tool if you’re pasting anything sensitive. Source: kivtools.com.

Update 30 September 2026: kivtools.com has not resolved since 25 September. The domain is still registered, but it currently points nowhere, so none of the above can be re-checked. If it comes back, this note goes; if it stays down, the section comes out.

Worth knowing: closed-source, so the local-processing claim isn’t independently checkable the way an open repo would be. For formatting a config file, fine. For anything with secrets in it, prefer something you can read the source of.

How tools get into this round-up: they arrive through our submission form, they have to be for people who build software, and I verify what I can before writing. Free editorial coverage is at our discretion with no timeline — the standards page sets out how we review and what money does and doesn’t buy. Nothing in this piece was paid for.

Verified 20 September 2026 against each project’s own repo, docs and pricing pages; source links added 23 September after HOL’s maintainer correctly pointed out the sections cited sources without linking them.