Everyone is a Dev. Find the best AI dev tools. Build in public. Join the community. Newsletter → https://nitter.cf/t.co/tihZRT9wmm
California, USA
Joined March 2025
- Tweets3.7K
- Following217
- Followers381
- Likes432
Pinned Tweet
204 AI dev tools launched on our site this week. These 14 are the ones we'd tell a friend about, and 13 of them are open source.
Meet Highlights. New list every Friday.
Here's the first edition, for the week ending Oct 30:
1. SPACE from Perplexity, an open-source sandbox that runs agent code in isolated Firecracker microVMs through an HTTP API.
2. TokenEconomizer from @microsoft Research, a plugin for Claude Code, Codex, Copilot CLI and Pi that learns from finished sessions to cut agent cost on a repo.
3. tstack from @theo, Claude Code and Codex skills for filing, reviewing, babysitting and merging PRs.
4. Brotli Checker from @trailofbits, an agent skill and scanner that finds Brotli decoder copies that may need patching for CVE-2026-111497.
5. @superplanehq, an open-source software factory that turns routine backlog issues into review-ready PRs.
6. Agent Governance Toolkit, also from Microsoft, which enforces policy, identity, audit and sandboxing on agent tool calls in deterministic code.
7. Expo Skills from @expo, official agent skills for building, deploying and debugging Expo and EAS apps.
8. SoL-Pi from NVIDIA NVlabs, a research extension for the Pi coding agent that cuts repeated turns and context replay.
9. DevBox from @RudderStack, a self-hosted harness that goes from ticket to plan, code, tests, self-review and a draft PR.
10. LayerFS, which gives each agent its own throwaway writable workspace over shared, deduplicated history in a single SQLite file.
11. Rystic, local stateful simulators of the Slack, GitHub, Stripe, Linear and Resend APIs for testing agents.
12. agentplay from @EntelligenceAI, which turns Claude Code, Codex, Cursor and Copilot transcripts into interactive maps and timelines, all locally.
13. Strands Harness SDK from AWS, a Python and TypeScript SDK for production agents.
14. @entirehq, a distributed Git forge that stores agent sessions and intent alongside the code.
The pattern this week is hard to miss. Most of the list is about keeping agents safe, cheap or reviewable rather than making them write more code: sandboxes, governance, cost cutting, transcript replay and PR babysitting. The tools that write code are getting good, and the work is moving to everything around them.
See you next Friday!
The scary part of putting a coding agent in CI is giving it write access to your repo. GitHub Agentic Workflows, out of preview this week, gets around that, and the design is worth knowing even if you never use it.
The agent never gets write access. It runs read-only in a sandboxed container behind a firewall, and it can only request the outputs you allowed in the workflow file, like opening an issue or a PR. A separate step checks each request, and a threat-detection job scans proposed changes before anything is applied.
You write the workflow in Markdown with a short header for the trigger and the allowed outputs. It compiles to normal Actions YAML on your existing runners and policies. The agent can be Copilot, Claude, Codex or Antigravity, and only the default Copilot engine needs a Copilot license. Run gh aw forecast first to estimate your monthly AI credit spend, including the safety scans.
GitHub's proof point is a quote from Accenture: one lead saved an estimated 29 hours a week, and his team's production errors fell from about 11,000 a day to under 1,000. That's one customer's report, not a benchmark. A detail I find more convincing is that gh-aw's own release notes are generated by one of its workflows.
If you try it, start with CI failure triage. It's mostly reading, the output is an issue instead of a code change, and you'll know within a week whether the agent's diagnoses are worth reading.
More on coding agents every week: EveryDev.ai/newsletter
Your model scores 72 on a benchmark, but you can't see which samples it got wrong. Lighteval from @huggingface lets you inspect results sample by sample, so you can debug failures instead of trusting one number. It runs 1000+ tasks across vLLM, SGLang, TGI and LiteLLM, supports custom tasks and metrics, and is open source under the MIT license.
New on EveryDev.ai: Leon
Open-source personal AI assistant built around tools, context, memory, and agentic execution that can run locally on.
Smart, controlled, and agent execution modes + Google Cloud, pricing, and setup details: @grenlouis everydev.ai/tools/leon
Building vector indexes on CPU gets slow once your embeddings hit the hundreds of millions. NVIDIA cuVS from @nvidia moves nearest neighbor search, vector compression and clustering onto the GPU, with CAGRA, IVF-PQ, IVF-Flat and ScaNN indexes. You can build on GPU and deploy on CPU. It's Apache-2.0, with APIs for C, C++, Python, Java, Rust and Go.
RAG answers fall apart when PDFs lose their tables, charts and page structure, and nobody can tell where a claim came from. Knowhere from @snapfill parses complex PDFs and PowerPoint into hierarchy-aware memory with page-grounded citations. It's open source and self-hostable with Docker, PostgreSQL, Redis and S3-compatible storage, and exposes retrieval over MCP for Cursor, VS Code, Claude and Codex.
AI coding agents are taking on the jobs that come after “the code works.”
Replit is investigating production failures. Mintlify is drafting documentation updates. VS Code is comparing agents on the same task, while Amp lets teams run agent environments on their own infrastructure.
That shifts the useful question from “which model writes the best code?” to “which parts of maintaining software can we actually hand over?”
There's progress on checking the answers, too. Microsoft's Scope tests how agents use software, and Anthropic is exploring formal proofs for open-source code.
More delegation, with different rules for review, permissions and cost.
This week's AI dev news digest looks at what's available, what's experimental, and where a person still makes the call. everydev.ai/p/news-weekly-ai…
Building slide decks by hand eats time you'd rather spend shipping. Guizang PPT Skill from @op7418 is an agent skill that generates single-file HTML decks, slide illustrations and social covers, with a built-in presenter mode for notes, timers and a dual-window audience screen. It works with Claude Code, Codex and Cursor, and it's open source under AGPL-3.0.
Stitching together prompts for lit review, experiments and paper drafts by hand gets old fast. AI Research Skills Library from @orch_research gives coding agents 98 skills across 23 categories, covering ideation, training, fine-tuning, evaluation, RAG, MLOps and LaTeX paper writing with citation checks. One npx command auto-detects your agent. MIT licensed, works with Claude Code, Cursor, Codex, Gemini CLI, OpenCode and Qwen Code.
Juggling separate gateways for your APIs and your LLM calls? Apache APISIX from @TheASF handles both. It's an open-source gateway built on NGINX and etcd, with hot-reloadable plugins, load balancing and circuit breaking. Its AI proxy adds fallback and load balancing across LLM providers, plus token-based rate limiting. It works with OpenAI, Anthropic, AWS Bedrock, DeepSeek and Ollama.
Agents that write to a shared knowledge store can corrupt it, and rolling back is painful. Omnigraph from @modernrelay is an open-source graph database on object storage with git-style branching and row-level three-way merge, so agents work on a branch. It fuses vector, full-text and graph search with rrf(), and runs on S3, R2, GCS or MinIO in open Lance format.
