Weekly AI Tools Radar — Issue #2 (June 22-29, 2026): Agent UI, Codebase Memory, AI Video & Agent Security

Most of this week's useful signal sat below the product surface: agent UIs, codebase memory, local search, video/voice tooling, and security checks for agent skills. GitHub Trending supplied most of the open-source candidates; Product Hunt added a few narrower end-user launches from June 27-28. I did not promote X-only launch posts unless they had a stable product page and pricing trail.

Fast scan

Tool	Cluster	Pricing	What makes it different	Try-it call
Persona.js	Agent UI	Free / open-source, per Product Hunt launch page 1	Framework-free, backend-agnostic, WebMCP-native chat widget for existing frontends 2	Try if you want a copilot inside an existing app without rebuilding around React.
codebase-memory-mcp	Codebase context	Free self-host, MIT license 3	Indexes a repo into a persistent knowledge graph; the project claims 158 languages, sub-ms queries, and 99% fewer tokens 3	Try on large repos where coding agents keep burning context on rediscovery.
Orca	Parallel coding agents	Free self-host, MIT license 4	Agent development environment for running a fleet of coding agents with your own subscriptions 4	Try if you already use more than one coding agent and need orchestration rather than another editor.
OpenMontage	Agentic video	Free self-host, AGPL-3.0 license 5	Open-source video production system with 12 pipelines, 52 tools, and 500+ agent skills 5	Try only if you can tolerate a heavy media stack and AGPL obligations.
Voicebox	Voice AI	Free self-host, MIT license 6	Open-source AI voice studio for cloning, dictation, and creation 6	Try for local voice experiments before paying for a hosted voice workflow.
Dotient	Local semantic search	$5 one-time Standard license, up to 3 computers, offline, all future updates included 7	Local-first visual/text semantic search for personal files; files stay on-device 7	Try if your pain is finding images, PDFs, and reference files, not enterprise document search.
Agent-Reach	Agent web access	Free self-host, MIT license 8	One CLI for agents to read and search Twitter/X, Reddit, YouTube, GitHub, Bilibili, Xiaohongshu, and more without API fees 8	Try for research agents that need public web context; review platform terms before automating heavily.
SkillSpector	Agent security	Free self-host, Apache-2.0 license 9	NVIDIA security scanner for AI agent skills, aimed at vulnerabilities and malicious patterns 9	Try if your team installs third-party agent skills or shares skill packs internally.

Cluster 1: agent surfaces are moving into the app

Persona.js

Persona.js is the cleanest Product Hunt launch in this week's developer-facing batch. Its launch page describes a lightweight, open-source AI chat UI library that embeds into any website, supports streaming chat, voice, theming, and WebMCP-native tool discovery 1. The official site backs that up with plain install paths: npm install @runtypelabs/persona, script-tag installs, SSE backends, Shadow DOM style isolation, and examples for AI SDK, OpenAI Agents, LangGraph.js, Hono, Express, and SvelteKit 2.

The differentiator is not another chat panel. It is the WebMCP angle: page actions can be exposed through document.modelContext, so the assistant can use tools already present in the host page with user approval 2. That matters for teams trying to add agentic behavior to existing software without creating a second backend API surface just for the assistant.

Pricing: Product Hunt lists Persona.js as free, and the project presents itself as open-source 1. Try it if you have an app where users already perform structured actions, such as searching a catalog, editing a calendar, or manipulating a workspace. Skip it if your assistant is just a help-doc bot; simpler chat widgets will be faster to ship.

Orca

Orca calls itself an ADE, an agent development environment, for working with a fleet of parallel agents. The repo says it can run coding agents with the user's own subscriptions and is available on desktop and mobile 4. GitHub Trending showed 2,769 stars in the weekly window, which is enough signal to treat it as more than a quiet repo update 10.

github.com · GitHub repository

stablyai/orca

https://github.com/stablyai/orca

Loading content card…

The practical question is whether you need coordination more than model access. If your workflow is one developer prompting one coding agent, Orca is probably premature. If you already split work across Claude Code, Codex-style tools, opencode, Cursor agents, and worktrees, the fleet metaphor starts to make sense. Pricing is effectively open-source self-hosting under MIT; model and agent subscriptions remain yours 4.

Cluster 2: codebase memory is becoming its own layer

codebase-memory-mcp

codebase-memory-mcp is a high-performance MCP server that indexes codebases into a persistent knowledge graph. The repo claims 158 supported languages, sub-ms queries, a single static binary, zero dependencies, and 99% fewer tokens 3. It also showed 8,926 weekly stars on GitHub Trending 10.

github.com · GitHub repository

DeusData/codebase-memory-mcp

https://github.com/DeusData/codebase-memory-mcp

Loading content card…

The differentiator is persistence. Many coding-agent workflows still re-scan the same files, ask the model to infer architecture repeatedly, and then lose that context at session end. A graph-backed MCP server is a better fit for long-lived repos, especially where the same services, symbols, and conventions recur across tasks.

Pricing: free self-host under MIT 3. Try it on a real monorepo, not a toy project. The evaluation should be simple: does it reduce repeated explanation and context loading enough to justify another local service?

Agent-Reach

Agent-Reach gives agents a single CLI for reading and searching public sources including Twitter/X, Reddit, YouTube, GitHub, Bilibili, and Xiaohongshu, with the repo positioning it as zero API fees 8. It gained 7,692 stars in the GitHub Trending weekly view 10.

github.com · GitHub repository

Panniantong/Agent-Reach

https://github.com/Panniantong/Agent-Reach

Loading content card…

The tool is interesting because agent research quality often fails before reasoning starts: the agent cannot reliably reach the places where useful context lives. Agent-Reach tries to package those reach capabilities as a local tool layer. That is useful, but it also means teams should read platform terms and rate-limit behavior before putting it inside an unattended workflow.

Pricing: free self-host under MIT 8. Try it for internal research agents or demos. Be more cautious for customer-facing automation.

Cluster 3: media tools are getting more agentic, but the stacks are heavier

OpenMontage

OpenMontage was the loudest GitHub signal this week: 18,703 weekly stars on GitHub Trending, ahead of every other AI-related repo I found in the weekly all-language and Python lists 10. The repo describes an open-source, agentic video production system with 12 pipelines, 52 tools, and 500+ agent skills 5.

github.com · GitHub repository

calesthio/OpenMontage

https://github.com/calesthio/OpenMontage

Loading content card…

The differentiator is scope. This is not a single text-to-video wrapper; it tries to turn a coding assistant into a video production studio spanning scripting, media generation, editing, audio, and rendering. That ambition is also the risk. Teams should expect dependency work, GPU/media-tool friction, and operational maintenance.

Pricing: free self-host under AGPL-3.0 5. Try it if you want to prototype an internal agentic video pipeline. Do not adopt it casually for commercial production without reviewing the license and deployment model.

Voicebox

Voicebox is an open-source AI voice studio for cloning, dictation, and creation. The GitHub repo lists MIT licensing, TypeScript, and topics around voice cloning, Whisper, CUDA, MLX, and Qwen3 TTS 6. It also had 3,883 weekly stars in GitHub Trending 10.

github.com · GitHub repository

jamiepine/voicebox

https://github.com/jamiepine/voicebox

Loading content card…

The useful thing here is control. Hosted voice tools are convenient, but they often hide model choice, storage, and consent flows. Voicebox is more interesting for teams that want a local or inspectable studio before they commit to a vendor. The caveat is the same as with any voice-cloning tool: only use it with consent and clear data handling.

Pricing: free self-host under MIT 6. Try it for prototyping voice workflows and evaluating local model quality. If you need production-grade compliance, the open-source starting point is not the whole answer.

Cluster 4: local-first search and agent security are pragmatic upgrades

Dotient

Dotient launched on Product Hunt on June 28 and describes itself as a local-first desktop app for organizing and searching personal files with ML-powered visual search 11. The official site is more specific: files stay on-device, the app uses quantized vision/text embeddings, and Standard pricing is a $5 one-time purchase for up to 3 computers with offline use and future updates included 7.

Dotient is not trying to be enterprise RAG. It is closer to a personal visual memory layer for photographers, designers, art directors, and people with chaotic camera rolls. The signal is the product shape: local-first, cheap, and narrow.

Try it if you have a reference-image or PDF retrieval problem on your own machine. Skip it if you need team permissions, shared indexes, or compliance controls.

SkillSpector

SkillSpector is NVIDIA's security scanner for AI agent skills. The repo says it detects vulnerabilities, malicious patterns, and security risks; it is Apache-2.0 licensed and was pushed during the current weekly window 9. It appeared in GitHub Trending Python with 2,312 weekly stars 12.

github.com · GitHub repository

NVIDIA/SkillSpector

https://github.com/NVIDIA/SkillSpector

Loading content card…

The differentiator is category timing. Agent skills are becoming portable code bundles, and many teams are starting to copy them the way they used to copy shell scripts. A scanner dedicated to that layer is a sensible addition to CI or review workflows.

Pricing: free self-host under Apache-2.0 9. Try it before importing third-party agent skills into internal environments. It is a guardrail, not a substitute for code review.

Quick picks

AWS Agent Toolkit for AWS: official AWS-supported MCP servers, skills, and plugins for agents building on AWS; Apache-2.0 licensed and recently pushed during the current weekly window 13. Try if your agent work already touches AWS APIs; skip if you are cloud-agnostic this quarter.
Folio AI: Product Hunt's June 27 leaderboard placed Folio AI at #1, with a pitch around PowerPoint and Google Slides 14. Official pricing starts with 500 free credits, then $0.03/credit on Starter, $10/month Plus, $49/month Pro, $199/month Enterprise, and $499/month Ultimate 15. Try if slides are a recurring bottleneck; skip if you only need occasional image-to-slide cleanup.
QApilot's CoWork: launched this week on Product Hunt as an AI-assisted mobile test automation product for existing test cases 16. The official site says it imports Jira, TestRail, spreadsheets, or other test-management assets, then runs on real iOS, Android, and Flutter devices with human-approved replanning 17. Pricing is not public on the official page; Product Hunt labels it payment required 16. Try if mobile QE backlog is the constraint.
discode.ai: Product Hunt ranked it #1 on June 28 and describes it as one interface for 100+ AI models 18. The official site emphasizes model routing, on-device sensitive-data handling before prompts leave the device, and answer checks across multiple models 19. Public pricing is not posted on the official site; Product Hunt lists free options 20. Try if EU-friendly routing and model-choice transparency matter more than raw API cost.
palmier-pro: open-source macOS video editor built for AI, GPL-3.0 licensed, with 5,034 weekly stars on GitHub Trending 10 21. Try if you want an inspectable macOS-native AI video editor; wait if you need cross-platform support.

What I would actually try first

For engineering teams, start with codebase-memory-mcp and SkillSpector. One tackles context waste; the other tackles a risk that will grow as skill packs spread. For product teams, Persona.js is the highest-leverage experiment because it lets you put an agent into an existing interface without betting the whole frontend on a new framework. For media teams, OpenMontage is the most ambitious, but Voicebox is the easier first test because the scope is narrower and the license is simpler.