Telexed

AI news through a solo-operator lens — only what changes your day22 of 412

FILTER[All][Agents & tools][Models & API][Generative media][Infra & SaaS][ASO & growth][Indie business][Idea signals][Other][★6+ high-signal]

GitHub Trending Weekly ✕clear filters

Mon, May 184 dispatches

#0412
#0412Agents & tools GitHub Trending Weekly2 days ago
`react-doctor` scores React code quality from the CLI
80radar
react-doctorReact diagnostics CLI — outputs a 0-100 quality score
A single npx run returns a 0-100 score plus concrete diagnostics. Useful as a cheap CI guardrail for agent-written React.
- npx -y react-doctor@latest . scans Next.js, Vite, and React Native projects with framework-aware rules.
- Checks cover state/effects, performance, architecture, security, accessibility, and dead code. That maps well to common agent regressions.
- The included GitHub Action can comment on PRs and expose a score output. Set fail-on to block low-quality changes.
- react-doctor.config.json supports rule, file, and override ignores. Use narrow overrides to keep coverage intact.
- install writes React best-practice guidance into Claude Code, Cursor, Codex, OpenCode, and 50+ agents.
Source: github.com/millionco/react-doctorRead original →
FIG-4121:1
80radar
FIG-4121:1
#0411
#0411Agents & tools GitHub Trending Weekly2 days ago
`Bun`: one JavaScript toolchain for runtime, bundling, tests, and packages
80radar
BunJavaScript toolchain — runtime and package manager in one
One executable covers node, package install, scripts, tests, and TS/JSX execution. Fewer moving parts make small app stacks faster to set up and easier to keep lean.
- bun run index.tsx supports TypeScript and JSX out of the box, cutting the usual transpiler setup for quick prototypes.
- The CLI includes bun test, bun install, bunx, and script running, so one binary can replace several Node.js dev dependencies.
- Built in Zig on JavaScriptCore, it targets lower startup time and memory use than typical Node-based tooling.
- Linux, macOS, Windows, Docker, npm, and Homebrew installs are covered; adoption friction is low for existing projects.
Source: github.com/oven-sh/bunRead original →
FIG-4111:1
80radar
FIG-4111:1
#0410
#0410Agents & tools GitHub Trending Weekly2 days ago
`Superpowers`: skill framework for coding agents
60radar
SuperpowersCoding-agent skills framework — enforces spec-to-TDD workflow
A reusable skills layer turns coding agents into spec-plan-TDD executors. Useful if your agent often jumps into code too early.
- Runs on Claude Code, Codex CLI, Codex App, Cursor, Copilot CLI, and others — not tied to one agent harness.
- Workflow is explicit: clarify intent, produce readable spec chunks, get design sign-off, then generate an implementation plan.
- The method pushes red/green TDD, YAGNI, and DRY; its value is guardrails, not a stronger base model.
- Subagent-driven execution can run for hours when the plan is tight. Best fit: scoped refactors, feature slices, test expansion.
Source: github.com/obra/superpowersRead original →
FIG-4101:1
60radar
FIG-4101:1
#0409
#0409Agents & tools GitHub Trending Weekly2 days ago
`Pyrefly`: Meta’s Fast Python Type Checker and Language Server
80radar
PyreflyPython type checker — fast checks plus IDE features
Python projects get fast type checks plus IDE features in one tool. Strong fit for typed backends and AI-heavy codebases where slow feedback kills iteration.
- Claims 1.85M+ lines/sec and checks PyTorch 15x faster than Mypy and Pyright; save-time IDE rechecks usually finish under 10ms.
- Runs as both CLI and language server, covering navigation, autocomplete, hover info, inlay hints, and semantic highlighting with consistent results.
- Built-in Pydantic and Django support reduces false friction for typical SaaS backends and API projects.
- Migration path is practical: pyrefly init, pyrefly suppress, and pyrefly infer let teams start with one file instead of a full rewrite.
Source: github.com/facebook/pyreflyRead original →
FIG-4091:1
80radar
FIG-4091:1

Wed, May 132 dispatches

#0408
#0408Agents & tools GitHub Trending Weekly7 days ago
`CloakBrowser`: stealth Chromium as a drop-in for Playwright/Puppeteer
50radar
CloakBrowserBrowser tool — patched Chromium for stealth automation
Uses a real Chromium build with source-level fingerprint patches instead of JS spoofing. Useful when browser automation gets blocked, but policy and abuse risk make it a selective tool, not a default stack choice.
- Claims 30/30 bot-detection tests passed, including Cloudflare Turnstile, FingerprintJS, and BrowserScan.
- Keeps the Playwright/Puppeteer workflow mostly intact: swap imports to launch() and keep the rest of the code nearly unchanged.
- Ships 49 C++ fingerprint patches across canvas, WebGL, audio, fonts, GPU, WebRTC, timing, and automation signals.
- humanize=True adds mouse, keyboard, and scroll behavior simulation; first run auto-downloads a ~200MB Chromium binary.
- Includes a self-hosted profile manager via docker run ... cloakbrowser-manager, positioning it against Multilogin-style setups.
Source: github.com/CloakHQ/CloakBrowserRead original →
FIG-4081:1
50radar
FIG-4081:1
#0407
#0407Agents & tools GitHub Trending Weekly7 days ago
`agentmemory`: persistent memory layer for coding agents
70radar
agentmemoryMemory layer — persistent context across agents
Adds a shared memory server across agent sessions, so setup choices and code context carry over automatically. Broad MCP/hook support plus strong retrieval benchmarks make it worth testing now.
- One server works across Claude Code, Cursor, Codex CLI, Gemini CLI, Cline and other MCP/HTTP clients, so context is no longer trapped in one tool.
- Retrieval claims are strong: 95.2% R@5 and 98.6% R@10 on LongMemEval-S, well above its BM25-only fallback.
- It stores implementation decisions like auth stack, file paths, and test coverage, then injects them into later sessions without manual re-explaining.
- v0.9.0 added agent-memory.dev, @agentmemory/fs-watcher, and tighter MCP/server syncing so hooks, viewer, and memory state stay aligned.
Source: github.com/rohitg00/agentmemoryRead original →
FIG-4071:1
70radar
FIG-4071:1

Mon, May 114 dispatches

#0406
#0406Agents & tools GitHub Trending Weeklylast week
`addyosmani/agent-skills`: reusable engineering skills for coding agents
50radar
agent-skillsSkill set — reusable workflows for coding agents
Turns senior-engineering workflows into reusable SKILL.md packages and lifecycle commands like /spec, /build, and /ship. Useful if you want stricter agent output without building your own prompting layer, but it's process scaffolding, not a model capability jump.
- Maps the dev lifecycle into 7 commands: /spec, /plan, /build, /test, /review, /code-simplify, /ship.
- Packages review gates, testing habits, and implementation workflows as portable skills, so agent behavior stays more consistent across tools.
- Supports multiple agent shells such as Claude Code, Cursor, Gemini CLI, Windsurf, and OpenCode, which lowers lock-in.
- The practical value is in standardizing execution quality; it won't make a weak agent smarter, but it can reduce sloppy output and missed steps.
Source: github.com/addyosmani/agent-skillsRead original →
FIG-4061:1
50radar
FIG-4061:1
#0405
#0405Agents & tools GitHub Trending Weeklylast week
`Flue`: headless TypeScript agent harness framework
60radar
FlueFramework — runs UI-less agent workflows in code
A fully programmable agent runtime that keeps Claude Code-style workflows but drops the human-in-the-loop UI. Promising if you want deployable agents across Node.js, Cloudflare, or CI, but the experimental API makes it an early bet.
- Runs agents in a default virtual sandbox via just-bash instead of full containers, so latency and cost can be much lower for high-traffic workloads.
- Puts most agent behavior in Markdown such as skills, context, and AGENTS.md, which reduces orchestration code and makes prompt-layer changes faster.
- Targets multiple runtimes including Node.js, Cloudflare, GitHub Actions, and GitLab CI/CD; useful when the same agent must ship across app backend and automation jobs.
- Typed outputs are built into the flow with schema validation in session.prompt(), which helps when wiring agents into production endpoints.
- Support-agent setup mounts knowledge from an R2 bucket directly into the filesystem and uses built-in search tools, a practical pattern for docs-backed support bots.
Source: github.com/withastro/flueRead original →
FIG-4051:1
60radar
FIG-4051:1
#0404
#0404Agents & tools GitHub Trending Weeklylast week
`UI-TARS Desktop` Adds Free Remote Browser and Computer Operators
70radar
UI-TARS DesktopDesktop agent — free remote browser and computer control
Open multimodal agents are getting usable outside closed IDEs. Free remote browser and computer control makes this worth testing for UI automation and ops work.
- The stack ships two tracks: Agent TARS for CLI/Web UI workflows and UI-TARS Desktop for native GUI control across desktop and browser.
- The 2025-06-12 update added free Remote Computer Operator and Remote Browser Operator, lowering the setup cost for agent-driven automation.
- Agent TARS CLI v0.3.0 added multi-tool streaming, timing stats, and an event stream viewer, which helps when debugging long agent runs.
- This is broader than a model demo: local/remote operators plus MCP tool integration make it more practical for repetitive QA, back-office, and browser tasks.
Source: github.com/bytedance/UI-TARS-desktopRead original →
FIG-4041:1
70radar
FIG-4041:1
#0403
#0403Agents & tools GitHub Trending Weeklylast week
`DeepSeek-TUI`: Terminal Coding Agent with Auto Model Routing
70radar
DeepSeek-TUITerminal coding agent — all-in-one TUI with rollback
Brings a full coding-agent stack to the terminal with streamed reasoning, approval gates, and workspace rollback. Strong if you want Codex/Claude Code-style flow on DeepSeek without building your own wrapper.
- Runs from the deepseek command and supports file edits, shell, git, web, sub-agents, and MCP in one TUI.
- --model auto also picks thinking level per turn, lowering manual model switching during mixed-cost tasks.
- Ships with 1M-token context handling, prefix-cache cost telemetry, and session resume for long jobs.
- /restore and side-git snapshots add rollback without touching the repo's own .git, which is safer for active projects.
Source: github.com/Hmbown/DeepSeek-TUIRead original →
FIG-4031:1
70radar
FIG-4031:1

Sun, May 101 dispatches

#0402
#0402Agents & tools GitHub Trending Weeklylast week
`Ouroboros`: spec-first Agent OS for replayable AI coding
60radar
OuroborosAgent OS — spec-first AI coding workflow
Adds a local-first runtime that turns agent coding into a replayable, policy-bound workflow. Worth testing if you want less prompt thrash and more verifiable runs across multiple coding CLIs.
- Uses a fixed loop of interview -> crystallize -> execute -> evaluate -> evolve, pushing requirements capture ahead of code generation.
- Targets multiple runtimes including Codex CLI, Claude Code, OpenCode, Hermes, Copilot CLI, and Kiro, reducing lock-in.
- The pitch is a 3-stage automated evaluation gate plus an immutable seed spec, so architecture drift and manual QA get constrained early.
- Setup is oriented around local tooling and MCP registration, with ooo interview "..." as the entry point instead of freeform prompting.
Source: github.com/Q00/ouroborosRead original →
FIG-4021:1
60radar
FIG-4021:1

Sat, May 91 dispatches

#0401
#0401Agents & tools GitHub Trending Weekly2 weeks ago
`Browserbase Skills` plugs Browserbase automation into `Claude Code`
60radar
Browserbase SkillsBrowser automation skill pack — unifies CLI, traces, cookie sync
A ready-made skill pack turns Claude Code into a browser automation agent with bb CLI, tracing, CAPTCHA handling, and UI testing. Useful if you want one install to cover QA, scraping, and auth-heavy browser tasks.
- The bundle ships 12 focused skills, including browser control, bb CLI workflows, tracing, anti-bot detection, cookie sync, and UI testing.
- Install flow is short: npx skills add browserbase/skills, then /plugin install browse@browserbase inside Claude Code.
- site-debugger and browser-trace stand out because they target flaky automation debugging, not just happy-path browsing.
- cookie-sync plus persistent Browserbase contexts makes logged-in site automation much more practical than stateless browser runs.
Source: github.com/browserbase/skillsRead original →
FIG-4011:1
60radar
FIG-4011:1

Fri, May 82 dispatches

#0400
#0400Agents & tools GitHub Trending Weekly2 weeks ago
`CocoIndex`: Incremental Context Engine for Long-Horizon Agents
60radar
CocoIndexContext engine — incremental indexing for fresh context
Keeps agent context continuously fresh by recomputing only changed data. Useful when your app ingests docs, code, or Slack and stale RAG is the bottleneck.
- Re-indexes only the delta, so changed files get re-embedded without rerunning the full pipeline.
- Targets codebases, PDFs, meeting notes, inboxes, Slack, and videos, which fits multi-source agent backends.
- The API is declarative Python with connectors like localfs and postgres; setup looks closer to ETL than agent orchestration.
- It also ships a coding-agent skill for v1 patterns, lowering the chance an AI IDE generates the wrong integration code.
Source: github.com/cocoindex-io/cocoindexRead original →
FIG-4001:1
60radar
FIG-4001:1
#0399
#0399Agents & tools GitHub Trending Weekly2 weeks ago
`OpenAI Symphony`: spec and reference impl for autonomous coding-run orchestration
50radar
OpenAI SymphonyCoding orchestration — connects task queues to PRs
Moves agent ops up a level: watch a work queue, spin isolated coding runs, and attach CI, review, and walkthrough artifacts before merge. Still an engineering preview, but the spec alone is useful if you want a homegrown task-to-PR pipeline.
- Targets work management, not live supervision: a board event can trigger isolated implementation runs that return CI status, PR feedback, complexity analysis, and videos.
- Two paths are offered: build from SPEC.md, or run the experimental Elixir reference in elixir/README.md. That lowers the barrier to cloning the pattern without waiting for a polished product.
- The repo explicitly says trusted environments only. This is not a drop-in autopilot for production repos yet.
- It assumes harness engineering is already in place, so the immediate value is highest for teams or solo operators who already enforce tests, CI, and review gates.
Source: github.com/openai/symphonyRead original →
FIG-3991:1
50radar
FIG-3991:1

Thu, May 78 dispatches

#0398
#0398Agents & tools GitHub Trending Weekly2 weeks ago
`jcode`, a coding-agent harness built for multi-session workflows
60radar
jcodeCoding agent harness — lightweight for multi-session runs
Optimized for many concurrent agent sessions with unusually low memory and near-instant startup. Worth a look if you care more about local orchestration efficiency than another chat wrapper.
- With local embeddings off, one session uses 27.8 MB PSS; even the full mode is 167.1 MB, below several agent CLIs listed.
- At 10 sessions, baseline memory is 117.0 MB and full jcode is 260.8 MB`; rivals in the table climb to 1.6-3.2 GB**, which matters for parallel runs.
- Time to first frame is 14.0 ms; the comparison table puts Codex CLI at 882.8 ms and GitHub Copilot CLI at 1518.6 ms.
- The pitch is not model quality but harness efficiency: multi-session operation, customizability, and low overhead as the main product surface.
Source: github.com/1jehuang/jcodeRead original →
FIG-3981:1
60radar
FIG-3981:1
#0397
#0397Agents & tools GitHub Trending Weekly2 weeks ago
`awesome-codex-skills`: Practical `Codex` Skill Library
50radar
awesome-codex-skillsSkill library — reusable Codex workflows and automation
Installable skills add task-specific execution rules and external actions to Codex. Useful as a starter kit for repeatable CLI/API automation.
- Skills install into $CODEX_HOME/skills and activate by metadata, so setup stays lightweight and reusable.
- The repo covers coding, writing, collaboration, and analysis tasks, making it more than a prompt snippet list.
- Each skill is a folder with SKILL.md instructions, which keeps workflows versionable and easy to adapt.
- Examples like multi-agent orchestration and app actions push Codex beyond code generation into operational automation.
Source: github.com/ComposioHQ/awesome-codex-skillsRead original →
FIG-3971:1
50radar
FIG-3971:1
#0396
#0396Agents & tools GitHub Trending Weekly2 weeks ago
`Ruflo`: multi-agent orchestration layer for `Claude Code`
60radar
RufloOrchestration layer — adds swarms and MCP to Claude Code
Adds a full coordination layer on top of Claude Code: swarms, memory, hooks, federation, and an MCP server. Useful if you want agent workflows without building your own orchestrator, but the full install is much heavier than a plugin-only trial.
- The full npx ruflo init path exposes 98 agents, 60+ commands, 30 skills, hooks, a daemon, and MCP-backed memory.
- Plugin installs stay light with slash commands and agent definitions only, but skip MCP features like memory_store and swarm_init.
- Positioning is clear: let coding agents coordinate across machines, sessions, and trust boundaries instead of acting as isolated copilots.
- Workspace footprint differs sharply by install path; the full setup writes .claude/, .claude-flow/, CLAUDE.md, helpers, and settings.
Source: github.com/ruvnet/rufloRead original →
FIG-3961:1
60radar
FIG-3961:1
#0395
#0395Agents & tools GitHub Trending Weekly2 weeks ago
`mattpocock/skills`, Composable Agent Skills Instead of Heavy Process Kits
50radar
mattpocock/skillsAgent skills pack — modular prompt-based workflows
Break agent workflows into small reusable skills, not rigid process layers. Easy to copy for prompt assets and task setup.
- Install is one command: npx skills@latest add mattpocock/skills, then run /setup-matt-pocock-skills for tracker, labels, and docs path.
- The design argues against GSD, BMAD, and Spec-Kit style process ownership; smaller prompt units are easier to debug and swap.
- /grill-me and /grill-with-docs target the biggest agent failure mode: bad requirement capture before code starts.
- The repo pushes shared language and lightweight conventions so agents waste fewer words and pick up project jargon faster.
Source: github.com/mattpocock/skillsRead original →
FIG-3951:1
50radar
FIG-3951:1
#0394
#0394Agents & tools GitHub Trending Weekly2 weeks ago
`Craft Agents` OSS: document-first multi-agent workspace with MCP/API setup
70radar
Craft AgentsAgent workspace — document-first multi-agent GUI
A GUI-first agent workspace aims to remove MCP/API setup friction by letting agents wire sources, skills, and credentials directly. Interesting if you want Claude Code-style power without living in the CLI.
- Uses the Claude Agent SDK and Pi SDK together, with a document-centric UI instead of a code-editor-first workflow.
- Claims source setup can be done in-chat: add Linear, Slack, paste MCP JSON, or hand over an OpenAPI spec and let the agent configure it.
- Local MCP support matters: stdio servers can run as subprocesses from npx, Python scripts, or local binaries without a restart loop.
- Apache 2.0 and “built with itself” make it a useful reference for anyone building an agent-native desktop product, not just using one.
Source: github.com/lukilabs/craft-agents-ossRead original →
FIG-3941:1
70radar
FIG-3941:1
#0393
#0393Agents & tools GitHub Trending Weekly2 weeks ago
`Zed`, a high-performance multiplayer code editor from the Atom and `Tree-sitter` team
40radar
ZedCode editor — built for speed and multiplayer coding
Built around speed and real-time collaboration, not just another local editor. Cross-platform availability is practical today, but no web client yet keeps it from being a full default for every workflow.
- Positioning is clear: a high-performance, multiplayer editor, which puts collaboration in the core product rather than as a plugin add-on.
- Install paths already cover macOS, Linux, and Windows via direct download or package managers, so trial cost is low across common dev setups.
- Web is still unavailable and tracked separately, which limits use on locked-down machines, quick handoff links, and browser-first workflows.
- The repo exposes build docs for all three desktop platforms, a useful signal that self-hosting, patching, or deep customization is realistic.
- License compliance is strict enough that CI can fail on dependency metadata, so contributors should expect more operational rigor than a casual side project.
Source: github.com/zed-industries/zedRead original →
FIG-3931:1
40radar
FIG-3931:1
#0392
#0392Agents & tools GitHub Trending Weekly2 weeks ago
`Warp` open-sources its agentic terminal dev environment
70radar
WarpTerminal IDE — agentic workflows with external CLI support
Built-in agents are only half the story: external CLI agents plug into the same terminal-first workflow. The repo is now inspectable, so evaluation gets easier; mixed licensing keeps commercial reuse nuanced.
- Claude Code, Codex, and Gemini CLI are positioned as first-class options, so existing agent workflows can move over without a full tool reset.
- build.warp.dev exposes live agent activity like issue triage, specs, code changes, and PR review. That makes the product legible, not just marketable.
- Licensing is split: warpui_core and warpui are MIT, while most of the repo is AGPL v3. Great for learning and contribution, stricter for product embedding.
- Local setup is straightforward with ./script/bootstrap, ./script/run, and ./script/presubmit, which lowers the barrier to testing patches or auditing architecture.
Source: github.com/warpdotdev/warpRead original →
FIG-3921:1
70radar
FIG-3921:1
#0391
#0391Agents & tools GitHub Trending Weekly2 weeks ago
`free-claude-code`: route `Claude Code` through third-party or local models
50radar
free-claude-codeProxy tool — routes Claude Code to external/local models
A local proxy keeps Claude Code working while swapping Anthropic for OpenRouter, DeepSeek, NVIDIA NIM, or local runtimes. Unofficial, but practical for cost testing and model routing.
- Acts as a drop-in Anthropic proxy, so Claude Code can talk to six backends without changing its client protocol.
- Supports per-tier routing with MODEL, MODEL_OPUS, MODEL_SONNET, and MODEL_HAIKU, which is useful for cost/perf splits.
- Exposes /v1/models, so the native Claude Code model picker still works through the proxy instead of breaking UX.
- Adds optional Discord/Telegram bots and voice transcription via local Whisper or NVIDIA NIM, extending it beyond terminal-only use.
Source: github.com/Alishahryar1/free-claude-codeRead original →
FIG-3911:1
50radar
FIG-3911:1