Telexed

AI news through a solo-operator lens — only what changes your day18 of 412

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

GeekNews ✕clear filters

Mon, May 183 dispatches

#0412
#0412Agents & tools GeekNewsyesterday
`Semble`, code search for agents using 98% fewer tokens than grep
70radar
SembleCode search library — returns relevant chunks to cut tokens
Agents get only the relevant code chunks instead of reading whole files. If the latency claim holds in real repos, this is worth testing for repo-aware coding workflows.
- Semble accepts natural-language and code queries, then returns only the snippets likely needed by the agent.
- The headline claim is 98% fewer tokens than grep plus file reads, directly cutting context spend.
- Average repository indexing/search is listed around 250ms, so it fits interactive agent loops if accurate.
- Best use case: wiring it into custom agents or CLI flows where rg finds too much and full-file reads waste context.
Source: news.hada.io/topic?id=29626Read original →
FIG-4121:1
70radar
FIG-4121:1
#0411
#0411Agents & tools GeekNewsyesterday
`GemmaCI`: Free PR Reviews via GitHub Actions and `Gemma`
40radar
GemmaCIOpen-source PR reviewer — runs Gemma in GitHub Actions
AI review runs inside GitHub Actions with Gemma. It targets private repos where hosted PR reviewers move to paid plans.
- GemmaCI is open source, so the review workflow can live in the repo instead of a paid SaaS seat.
- The stated replacement target is CodeRabbit; the pain is private-repo PR review cost, not code quality benchmarking.
- No usage numbers or review-quality metrics yet. Treat it as a cost-saving experiment, not a proven reviewer.
Source: news.hada.io/topic?id=29618Read original →
FIG-4111:1
40radar
FIG-4111:1
#0410
#0410Agents & tools GeekNews2 days ago
`dev3000`: unified web app debugging timeline capture for AI agents
70radar
dev3000Web debugging CLI — unifies logs, console, and screenshots
Server logs, console output, network calls, screenshots, and user actions land in one timestamped trace. It gives coding agents better bug context than pasted logs.
- Captures dev server logs, browser console, network requests, screenshots, and interactions into one timestamped log.
- Uses Chrome DevTools Protocol to monitor the browser, so it fits normal web app debugging flows.
- Stores captured data under ~/.d3k/{project...}, making traces reusable across prompts and later investigations.
Source: news.hada.io/topic?id=29607Read original →
FIG-4101:1
70radar
FIG-4101:1

Sun, May 172 dispatches

#0409
#0409Agents & tools GeekNews2 days ago
`glowed`: Terminal Markdown Browser/Editor for `Ghostty`
40radar
glowedTerminal Markdown TUI — search and edit inside Ghostty
Keeps project docs searchable, previewable, and editable inside the terminal. The useful angle is the handoff from .md notes to an LLM CLI, worth trying if docs already live in-repo.
- glowed targets Ghostty and works as a TUI for finding, previewing, and editing Markdown files inside a project.
- The workflow is terminal-first knowledge management: inspect .md files without opening an IDE or separate notes app.
- LLM CLI handoff is the practical hook. Existing specs, notes, and prompts can become local context faster.
Source: news.hada.io/topic?id=29592Read original →
FIG-4091:1
40radar
FIG-4091:1
#0408
#0408Agents & tools GeekNews2 days ago
`Zerostack`: Unix-inspired coding agent written in pure Rust
50radar
ZerostackRust coding agent — minimal stack at roughly 7k LoC
A tiny agent stack with file ops, gated Bash, MCP, and multiple LLM providers. Worth watching if you want inspectable agent plumbing, not a daily-driver replacement yet.
- Written in Rust at roughly 7k LoC; small enough to audit or fork for custom agent experiments.
- Supports multiple LLM providers plus custom providers, so it is not locked to one model API.
- Tool surface covers file read/write/edit, grep, find, directory listing, permission-gated Bash, MCP, and Exa web tools.
Source: news.hada.io/topic?id=29584Read original →
FIG-4081:1
50radar
FIG-4081:1

Sat, May 162 dispatches

#0407
#0407Agents & tools GeekNews3 days ago
How `Claude Code` Works on Large Codebases: Best Practices and Starting Points
50radar
Large repos are read live through grep, file traversal, and reference chasing instead of prebuilt indexes. The real leverage is the harness around the model: CLAUDE.md, hooks, skills, plugins, and MCP setup order.
- It reads the local codebase directly instead of uploading or pre-indexing files, which keeps iteration close to the working tree.
- Repository performance depends on more than the model; CLAUDE.md, hooks, skills, plugins, and MCP wiring shape output quality.
- Setup order matters. A weak harness wastes tokens and context even when the underlying model is strong.
Source: news.hada.io/topic?id=29556Read original →
FIG-4071:1
50radar
FIG-4071:1
#0406
#0406Agents & tools GeekNews4 days ago
`git-sync`: mirror Git refs between remotes without a local checkout
40radar
git-syncGit CLI — mirrors remotes without local checkout
Streams refs and pack data straight from source to target, skipping clone and disk checkout. Useful for repo migration or backup pipelines when storage and transfer hops matter; everyday workflows stay mostly unchanged.
- upload-pack data is relayed directly into receive-pack, so objects avoid an intermediate local repo and disk write.
- Best fit is migration, mirroring, and backup automation for large repositories where clone time and storage are wasted overhead.
- Because it operates at ref/object transfer level, it solves infra plumbing rather than day-to-day feature development.
- It is a focused CLI, not a broader Git host replacement, so impact is practical but narrow.
Source: news.hada.io/topic?id=29547Read original →
FIG-4061:1
40radar
FIG-4061:1

Fri, May 154 dispatches

#0405
#0405Agents & tools GeekNews5 days ago
`Claude Code`/`Codex` skill adds post-task learning drills
40radar
Turns agentic coding into guided practice, not just output. After refactors or schema work, it offers 10-15 minute optional drills, so the tool can train judgment while shipping.
- After tasks like new file creation, schema changes, and refactors, it suggests a short follow-up exercise instead of ending at code output.
- The 10-15 minute format is small enough to fit real workflow, which makes deliberate practice more likely to stick.
- Useful signal for agent UX: copilots can improve operator skill, not only automate execution.
- Best read as a reusable interaction pattern for coding agents, not a platform-level product launch.
Source: news.hada.io/topic?id=29531Read original →
FIG-4051:1
40radar
FIG-4051:1
#0404
#0404Agents & tools GeekNews5 days ago
`DeerFlow` `2.0`, ByteDance’s Long-Running SuperAgent Harness
60radar
DeerFlowOpen-source agent harness — parallel orchestration for long tasks
Breaks multi-hour work into parallel sub-agents for research, coding, and content generation. Worth tracking if you want a self-hosted orchestration layer instead of bolting long-running loops onto a coding agent.
- Handles research, coding, and content jobs in one harness, aiming at workflows that run for minutes to hours rather than single prompts.
- v2 is a full rewrite with no shared code from the Deep Research-style v1, so expect architecture changes instead of an incremental patch.
- Built on LangGraph and LangChain, which lowers adoption cost for teams already using that stack but adds framework dependency.
- The useful angle is orchestration, not model novelty. Strong fit when one agent is no longer enough and ad-hoc scripts start to sprawl.
Source: news.hada.io/topic?id=29518Read original →
FIG-4041:1
60radar
FIG-4041:1
#0403
#0403Agents & tools GeekNews5 days ago
`Codex` Lands in the `ChatGPT` Mobile App
80radar
Active coding sessions now carry over to the phone with thread state, approvals, plugins, and project context intact. This makes long-running agent work less desktop-bound and worth trying for on-call fixes or quick intervention.
- The app pulls in active threads from laptop, devbox, and remote environments, so work can continue across devices without restarting context.
- Approvals, plugins, project context, and live status are exposed on mobile; that turns the phone into a control surface, not just a viewer.
- Fast check-ins become practical when an agent is mid-run and needs a small decision, reducing the need to sit in front of the main machine.
- The value is highest for long-running coding tasks and remote sessions; short local edits still feel faster on desktop.
Source: news.hada.io/topic?id=29516Read original →
FIG-4031:1
80radar
FIG-4031:1
#0402
#0402Agents & tools GeekNews5 days ago
`Codex` lands in the `ChatGPT` mobile app for remote agent control
80radar
CodexCoding agent — runs and reviews tasks remotely
You can now approve, monitor, and manage coding runs from your phone. Strong quality-of-life upgrade if Codex is already part of your workflow.
- Mobile support centers on remote control: approve actions, check progress, and manage runs away from the desktop.
- This extends existing agent sessions instead of replacing desktop use, so handoff during commute or downtime gets much smoother.
- The practical gain is faster unblock cycles when long-running tasks need a quick human approval.
- Teams are not the story here; solo operators get the clearest benefit from being able to keep jobs moving anywhere.
Source: news.hada.io/topic?id=29514Read original →
FIG-4021:1
80radar
FIG-4021:1

Thu, May 141 dispatches

#0401
#0401Agents & tools GeekNews6 days ago
`Ruflo`: Multi-Agent Orchestration for `Claude Code`
50radar
RufloPlatform — orchestrates 100+ specialized agents
Turns Claude Code into a swarm runner with 100+ specialized agents, memory, and cross-machine coordination. Interesting if you want heavier automation without building your own orchestration layer, but the pitch is still ahead of proof.
- Rebranded from Claude Flow, so this is packaging and positioning plus an expanded orchestration story.
- Setup is a single npx ruvflo init, lowering the barrier to trying swarm-style agent workflows fast.
- Claims include task-level learning, session memory, and self-organizing agents; that points to longer-running automation rather than one-shot coding help.
- Cross-machine coordination is the differentiator if real. Useful for parallel codegen or ops tasks, but reliability matters more than the demo surface.
Source: news.hada.io/topic?id=29480Read original →
FIG-4011:1
50radar
FIG-4011:1

Wed, May 132 dispatches

#0400
#0400Agents & tools GeekNews6 days ago
`Gram` `2.0.0` ships saner editor defaults
40radar
GramCode editor — streamlined defaults for a cleaner setup
Default behavior was cleaned up, but the reset comes with breaking changes. If you rely on prior editor assumptions, pin and verify before upgrading.
- Relative line numbers are now on by default, removing one of the editor's earlier odd defaults.
- Icons are now off by default, so the base UI is less noisy and more predictable.
- The release was bumped to a major version because these default changes can break compatibility with existing setups.
- Automatic language-server updates are now off by default, giving tighter control over tooling drift like ESLint changes.
Source: news.hada.io/topic?id=29464Read original →
FIG-4001:1
40radar
FIG-4001:1
#0399
#0399Agents & tools GeekNews7 days ago
`DeepSeek-TUI`: terminal coding agent for `DeepSeek` models
50radar
DeepSeek-TUICoding agent TUI — auto-selects DeepSeek models
Puts editing, shell, Git, web search, and sub-agents into one keyboard-first terminal loop. --model auto can switch pro/flash and reasoning per turn, so a DeepSeek-centric setup looks practical to test now.
- Covers file edits, shell commands, Git ops, web search, and sub-agent coordination in one TUI; fewer context switches than split CLI stacks.
- --model auto chooses between deepseek-v4-pro and flash plus reasoning levels like off/high/max; that adds built-in cost and latency control.
- The value is strongest for teams already standardized on DeepSeek; otherwise it is another agent surface, not a workflow reset.
Source: news.hada.io/topic?id=29447Read original →
FIG-3991:1
50radar
FIG-3991:1

Tue, May 122 dispatches

#0398
#0398Agents & tools GeekNewslast week
`Claude Code` unveils `Agent View` for managing multiple agents in one screen
80radar
Claude CodeCoding tool — makes multi-agent parallel work easier
You can launch, watch, and reply to background coding sessions from a single pane. The bigger win is persistence: a supervisor keeps agents alive without open terminals, which makes parallel task orchestration much more practical.
- claude agents becomes a control surface for spawning and monitoring multiple Claude Code sessions at once, reducing tab and terminal churn.
- Sessions stay alive without an attached terminal because a separate supervisor process owns them; long-running jobs become less fragile.
- This shifts Claude Code closer to a lightweight multi-agent orchestrator, useful for parallel bugfixes, research, and refactors from one screen.
Source: news.hada.io/topic?id=29431Read original →
FIG-3981:1
80radar
FIG-3981:1
#0397
#0397Agents & tools GeekNewslast week
`Claude Code` adds `/goal` for autonomous multi-turn execution
80radar
Claude CodeCoding agent — autonomous turns until goal completion
One prompt can now keep running until the task is done. A fast-model judge closes the loop each turn, making longer coding runs practical without manual babysitting.
- /goal keeps Claude Code running across multiple turns until completion, not just a single response cycle.
- A fast model evaluates goal completion after each turn; if unfinished, the next turn starts automatically.
- The loop removes manual re-prompting for longer tasks, but weak goal definitions can still waste time and tokens.
Source: news.hada.io/topic?id=29428Read original →
FIG-3971:1
80radar
FIG-3971:1

Mon, May 112 dispatches

#0396
#0396Agents & tools GeekNewslast week
`Open Design`, a local-first open-source alternative to `Claude Design`
50radar
Open DesignDesign tool — turns existing coding-agent CLIs into its engine
Instead of bundling its own agent, it turns whatever coding-agent CLI is already on your PATH into a design engine. That keeps switching costs low and makes it worth a quick test if your stack already includes Claude Code or Codex.
- It auto-detects 16 coding-agent CLIs on PATH, including Claude Code and Codex, so adoption starts from tools already installed.
- Ships with 31 built-in skills covering web prototypes, SaaS landing pages, dashboards, mobile apps, posters, and even PM docs like OKR or invoices.
- The local-first shape matters: design generation can ride existing CLI workflows instead of locking teams into one hosted agent product.
- Positioning is practical rather than novel model tech. Value comes from orchestration and reusable templates, not from a new foundation model.
Source: news.hada.io/topic?id=29376Read original →
FIG-3961:1
50radar
FIG-3961:1
#0395
#0395Agents & tools GeekNewslast week
Anthropic open-sources finance-focused AI agents, skills, and connectors
40radar
A reference repo lets the same finance workflow run via a Claude Cowork plugin or the Managed Agents API. Reusable agent patterns and connectors are the real takeaway; the vertical is narrow, but the orchestration shape is worth copying.
- The repo exposes two execution paths for one agent: Claude Cowork plugin and Managed Agents API, which hints at a portable agent architecture.
- It covers 10 workflows across investment banking, equity research, private equity, and wealth management, so the value is in packaged task design more than model novelty.
- Finance is not the main story here. Skills, connectors, and workflow wiring can be lifted into other document-heavy, approval-heavy SaaS niches.
- This is more useful as a reference implementation than a product launch. Good source material if you're building vertical agents with repeatable ops patterns.
Source: news.hada.io/topic?id=29372Read original →
FIG-3951:1
40radar
FIG-3951:1

`Semble`, code search for agents using 98% fewer tokens than grep

`GemmaCI`: Free PR Reviews via GitHub Actions and `Gemma`

`dev3000`: unified web app debugging timeline capture for AI agents

`glowed`: Terminal Markdown Browser/Editor for `Ghostty`

`Zerostack`: Unix-inspired coding agent written in pure Rust

How `Claude Code` Works on Large Codebases: Best Practices and Starting Points

`git-sync`: mirror Git refs between remotes without a local checkout

`Claude Code`/`Codex` skill adds post-task learning drills

`DeerFlow` `2.0`, ByteDance’s Long-Running SuperAgent Harness

`Codex` Lands in the `ChatGPT` Mobile App

`Codex` lands in the `ChatGPT` mobile app for remote agent control

`Ruflo`: Multi-Agent Orchestration for `Claude Code`

`Gram` `2.0.0` ships saner editor defaults

`DeepSeek-TUI`: terminal coding agent for `DeepSeek` models

`Claude Code` unveils `Agent View` for managing multiple agents in one screen

`Claude Code` adds `/goal` for autonomous multi-turn execution

`Open Design`, a local-first open-source alternative to `Claude Design`

Anthropic open-sources finance-focused AI agents, skills, and connectors