Open Source Radar — September 3: hands, voice, and models
Today's daily board is all about the gear bolted onto an agent that already writes code: a real browser it can drive and debug, a skill that scrubs the machine-tone out of its prose, a forecasting model that reads time series, and one cluster that serves every model the agent calls. Four fresh picks, none of them repeats from this week's earlier radars.
chrome-devtools-mcp (TypeScript, ~50,600 stars) — Google's official Chrome DevTools server for coding agents, and the clearest sign yet that the browser is becoming the agent's hands. It plugs into any MCP-capable agent and hands it the actual DevTools toolkit: record a performance trace and get actionable insights back, inspect network requests, take screenshots, and read console errors with source-mapped stack traces that point at your source rather than a bundled blob. Automation runs through Puppeteer and waits for each action to land, which is the difference between an agent that clicks hopefully and one that knows the click worked. Version 1.8.0 shipped on August 25, and there's a slim mode if you only need basic navigation. Two things to know before you wire it up: usage telemetry is on by default, and the performance tools may send trace URLs to Google's CrUX API for real-user data — both are opt-out. Apache-2.0.
humanizer (Python, ~40,400 stars) — The counter-programming to everything else on this list: an agent skill whose whole job is undoing the sound of AI writing. It works from a 35-pattern checklist drawn from Wikipedia's "Signs of AI writing," maintained by WikiProject AI Cleanup — the inflated-importance framing, the forced groups of three, the "not just X, it's Y" pivots, the chatbot sign-offs, the em dashes and emojis. The rewrite happens in two passes: a first draft that ignores the original structure, then a check of that draft against the patterns and against the source's actual claims, with a short critique shown alongside so you can see what still sounds synthetic. It refuses to invent facts — names, numbers, dates, quotes, and citations have to come from you or your sources — and it can match a writing sample you supply instead of its own defaults. Point it at a file and it edits only the prose, leaving code, data, frontmatter, and link targets untouched. It's plain Markdown, so any agent that supports skills can load it. Forty thousand stars for teaching agents to sound less like agents is the most telling number on today's board.
TimesFM (Python, ~29,800 stars) — Google Research's time-series foundation model, back on trending after TimesFM 3.0 landed in August with native multivariate forecasting and covariate support, covering both past-only and past-and-future variables without per-task tuning. The headline claim is competitive rather than architectural: the authors report rank-one results on fev-bench across 100 real-world forecasting tasks, on the TIME Benchmark across 50 domains and 98 evaluation tasks, and among foundation models on GIFT-Eval. It already sits inside BigQuery ML for SQL-scale forecasting and Google Sheets for spreadsheet users, so this is research that ships. The catch worth reading twice: the code is Apache-2.0 and weights through 2.5 stay permissive, but the 3.0 pretrained weights carry a separate non-commercial license, so production use of the default checkpoint is off the table for now. Reach for it when you need demand, traffic, or revenue forecasts across many series and don't want to train and babysit a model per line.
SIE (Python, ~3,050 stars) — The smallest project here and the one that best describes where the stack is going. Superlinked's inference engine is a self-hosted cluster that runs every model your agent calls through a single OpenAI-compatible API, replacing the usual pile of one model server per task. It covers search and retrieval, turning PDFs and Office files and scans into clean markdown, schema-valid structured output, content safety verdicts, and the agent loop itself, drawing on a catalog of a hundred-plus models that load on demand and get evicted under memory pressure. It ships Kubernetes and Helm configs, KEDA autoscaling, and Grafana dashboards, and it already speaks to LangChain, LlamaIndex, Haystack, DSPy, CrewAI, and the major vector stores. Apache-2.0. If your agent stack is currently four services and a scheduling spreadsheet, this is the shape it wants to become.
Worth watching this week: agents are acquiring senses and standards — a browser to check their own work in, a style guide to hide their fingerprints, and one inference tier underneath it all. The interesting question is no longer what the agent can do, but whether you can see it doing it.
Which of these would you wire into your agent first — the browser, the prose filter, the forecaster, or the model cluster? Tell us in the comments.
Sources: chrome-devtools-mcp (GitHub) · Chrome DevTools MCP tool reference · humanizer (GitHub) · Wikipedia: Signs of AI writing · TimesFM (GitHub) · TimesFM paper (arXiv) · SIE (GitHub) · Superlinked SIE docs
