New AI Tools Weekly #5: Agent infrastructure moves below the chatbot
This week’s strongest AI-tool signal is the operating layer below agents: codebase memory, production harnesses, parallel coding workspaces, media-skill systems, observability, and verifiable tool-call receipts.
The week’s strongest signal was not another chatbot UI. It was infrastructure around agents: memory for codebases, harnesses that turn agents into configurable services, workspaces for running multiple coding agents at once, and receipts for proving what tools an agent actually called.
Coverage note: Product Hunt’s weekly leaderboard remained inaccessible through direct fetches, so this issue used search-indexed Product Hunt candidates only as discovery leads. Entries made the cut only when a GitHub repo, official page, official blog, or X permalink could verify the product or launch detail. GitHub Trending and X supplied the clearest primary signals this week.
At-a-glance shortlist
| Theme | Tool | This week’s signal | Pricing / access | Try-it recommendation |
|---|---|---|---|---|
| Code context and governed data | Codebase Memory MCP | +6,372 weekly stars in the all-language GitHub Trending scan; repo at 10.8k stars | Open-source; MIT license shown in repo | Try if your coding agent burns tokens rediscovering architecture every session. 12 |
| Code context and governed data | Basedash groups and access controls | June 20 release adds group-level access and per-group AI context | Commercial BI product; signup available | Try if AI analytics needs different permissions and language for data teams, executives, and clients. 3 |
| Code context and governed data | NocoBase | +294 weekly stars in TypeScript Trending; repo positions itself as AI + no-code for business systems | Open-source repo; commercial ecosystem likely around deployments/services | Try if you want coding agents and human no-code builders working on the same internal app. 45 |
| Production agent harnesses | Amazon Bedrock AgentCore harness | GA blog published June 18; Amazon says CreateHarness + InvokeHarness can define and run an agent with managed memory, tools, skills, filesystem, and observability | AWS managed service; usage depends on Bedrock/AgentCore resources | Try if you already run on AWS and need isolation, memory, identity, and traces without writing the harness yourself. 6 |
| Production agent harnesses | Flue | +1,272 weekly stars in TypeScript Trending; programmable TypeScript harness for agents, workflows, sandboxes, skills, subagents, MCP, and observability | Open-source repo; Apache-2.0 license shown | Try if you want a code-first agent harness outside a single cloud vendor. 47 |
| Production agent harnesses | Google agents-cli | +182 weekly stars in Python Trending; CLI and skills for building, evaluating, deploying, and publishing agents on Google Cloud | Apache-2.0 repo; Google Cloud required for deployment features | Try if your coding assistant needs to scaffold ADK/Gemini Enterprise agents rather than hand you docs. 89 |
| Fleet coding workspaces | KiloCode | +3,674 weekly stars in TypeScript Trending; repo describes an all-in-one agentic engineering platform for VS Code, JetBrains, and CLI | Open-source repo; check license and hosted options before rollout | Try if you want a coding-agent environment rather than a single chat panel. 410 |
| Fleet coding workspaces | Orca | +997 weekly stars in TypeScript Trending; desktop/mobile workspace for running Codex, Claude Code, OpenCode, and other CLI agents side by side in separate worktrees | Open-source repo; desktop app distribution | Try if you want to compare multiple agents on the same task and merge the best result. 411 |
| Media and design production | OpenMontage | +2,867 weekly stars in Python Trending; repo claims 12 pipelines, 52 tools, and 500+ agent skills for video production | Open-source; AGPLv3 license shown | Try if you want reproducible AI video workflows with cost estimates and real footage paths. 812 |
| Media and design production | UI Skills | +506 weekly stars in TypeScript Trending; repo packages task-routed UI skill sets for design engineers | Open-source; MIT license shown | Try if your design/code agent keeps producing generic UI and needs narrower design instructions. 413 |
| Media and design production | Photoroom AI Ironing | Photoroom CEO announced AI Ironing on June 22; it removes garment wrinkles while preserving logo, texture, and stitching, and is available in AI Tools plus API for Plus customers | App feature; API access tied to Plus plan in the launch post | Try if ecommerce image teams spend real time retouching apparel wrinkles. 14 |
| Agent observability and proof | Foglamp | Official page positions it as observability for AI agents built on Vercel AI SDK, covering cost, latency, traces, evals, alerts, and per-agent spend | Start-free SaaS page | Try if agents are already in production and you need cost/regression alerts before users complain. 15 |
| Agent observability and proof | Fetch.ai AEVS | June 15 launch: signed, tamper-evident receipts for agent tool calls; secondary coverage describes HMAC-signed, hash-chained receipts and an off-chain design | Open-source SDK per coverage; Product Hunt launch | Try if you need portable evidence of tool calls for audits, payments, refunds, or disputes. 1617 |
Theme 1: Code context is turning into a product layer
Codebase Memory MCP was the cleanest GitHub signal of the week. It is not trying to be another coding agent. It gives agents a fast local knowledge graph of a repository: functions, classes, call chains, routes, and cross-service links. The README claims a Linux-kernel-scale index in 3 minutes, 158 languages through tree-sitter, hybrid LSP resolution for major languages, and large token savings versus file-by-file exploration. 2
The differentiation is straightforward: Cursor-style repo indexing, grep, and embeddings help find files; Codebase Memory is aiming at structural questions. If the repo is large enough that your agent keeps asking, "where is this defined?" or "what calls this endpoint?", this is the kind of layer that should sit under the agent.
Basedash and NocoBase show the same pattern outside pure code. Basedash’s June 20 release gives different groups access to different data sources, dashboards, chats, automations, MCP servers, and AI context. That matters because the moment an AI analyst sits on company data, governance becomes part of the product, not an admin afterthought. 3
NocoBase is the broader business-app version. Its README describes an AI + no-code platform where coding agents can handle setup, development, migration, and releases while people keep a WYSIWYG interface for data models, pages, workflows, and permissions. 5 The practical read: internal tools are moving toward mixed construction sites, where agents edit code and humans configure the app visually.
Theme 2: The agent harness is becoming a real deployment primitive
Amazon’s AgentCore harness GA is the enterprise marker. The blog says the harness wraps model choice, tools, skills, memory, identity, filesystem, browser, code interpreter, and observability behind CreateHarness and InvokeHarness. It also supports switching model providers mid-session, including Bedrock models, direct OpenAI, Gemini, and LiteLLM-supported providers. 6
Flue is the open, TypeScript-native counterweight. Its README frames it as a programmable harness for autonomous agents and workflows with sessions, tools, skills, instructions, filesystem access, and sandboxes. It also lists durable execution, subagents, MCP tools, observability integrations, and event channels such as Slack, Teams, Discord, and GitHub. 7
Google agents-cli attacks a narrower but practical problem: making a coding assistant competent at Google Cloud agent work. It installs CLI commands and skills for scaffolding, ADK code patterns, evals, deployment, publishing to Gemini Enterprise, and observability. The repo is explicit that agents-cli is not a coding agent; it is a tool and skill layer for the coding agent you already use. 9
The try-it split is clear. Use AgentCore if you want AWS-managed primitives and already live in that account. Try Flue if you want a framework you can read and deploy across runtimes. Try agents-cli if the friction is not agent architecture in general, but getting ADK/Gemini Enterprise agents built and shipped correctly.
Theme 3: Coding agents are getting workstations, not just chat boxes
KiloCode and Orca both point to a useful shift: the coding-agent interface is becoming a workspace for managing work, not just a prompt field.
KiloCode describes itself as an open-source coding agent for VS Code, JetBrains, and the CLI. The repo presents it as an all-in-one agentic engineering platform for building, shipping, and iterating faster with an open-source coding agent. 10 That makes it closer to a daily development surface than a narrow autocomplete tool.
Orca is more opinionated about orchestration. It runs CLI agents such as Codex, Claude Code, OpenCode, and others side by side, each in its own worktree. The README emphasizes mobile monitoring, parallel worktrees, terminal splits, design-mode browser capture, GitHub/Linear workflows, SSH worktrees, and commenting on AI diffs. 11
The differentiation against established IDE assistants is not model quality. It is process control. If you want one agent to edit a file, the default IDE assistant is enough. If you want five agents to attempt the same migration in separate worktrees, compare diffs, annotate the winner, and keep an eye on jobs from a phone, Orca is closer to the job.
Theme 4: Media and design tools are adopting the same agent-skill pattern
OpenMontage is the week’s strongest media-production repo. It turns an AI coding assistant into a video production studio: research, scripting, asset generation, editing, and final composition. The README distinguishes still-image animation from a real-footage path where the agent builds a corpus from free stock footage and open archives, retrieves motion clips, edits a timeline, and renders a finished piece. 12
The useful part is not the "AI makes video" pitch. It is reproducibility. OpenMontage shows example videos with prompts, pipelines, tools, and costs. For teams experimenting with AI video, that is more valuable than a glossy demo with no path back to the settings that produced it.
UI Skills brings the same idea to design engineering. Its CLI can route an agent through task-specific UI skill sets, and the repo’s purpose is plain: skills for design engineers. 13 This category keeps appearing because generic coding agents often produce generic interfaces. A smaller skill library can be more useful than a bigger model when the failure mode is taste and task framing.
Photoroom AI Ironing is a different kind of media tool: narrow, commercial, and practical. The launch post says it removes apparel wrinkles while preserving the garment’s logo, texture, and stitching, and that it is live in AI Tools plus available via API for Plus plan customers using
ironing.mode. 14 That is exactly the kind of domain-specific image model that should not be replaced by a general image editor prompt.Theme 5: Observability and proof are catching up to autonomous agents
Foglamp and AEVS are small compared with the big coding-agent repos, but they answer the same uncomfortable question: once agents can act, how do you know what happened?
Foglamp is observability for agents built on the Vercel AI SDK. Its page says the SDK captures cost, latency, tokens, distributed traces, evals, alerts, and per-agent spend for
generateText and streamText calls. 15 The best-fit user is not someone prototyping a toy agent. It is a team with enough production traffic that a cost regression, hallucinated policy, or bad answer can become a support problem.AEVS goes after proof rather than monitoring. Fetch.ai’s launch post says AEVS is now live on Product Hunt as an Agent Execution Verification System, and independent coverage says it records tool name, inputs, outputs, timing, status, duration, and sequence position, then produces HMAC-signed, hash-chained receipts. 1617
The key limitation is just as important: the same coverage describes AEVS as tamper-evident, not tamper-proof, and off-chain rather than blockchain-anchored. 16 That makes it easier to adopt than heavier verifiable-compute systems, but teams should treat it as an audit trail, not a cryptographic guarantee that the underlying execution could not be manipulated.
What to try first
For builders short on time, the practical order is:
- If your problem is agent context: try Codebase Memory MCP before switching coding tools. It targets the retrieval layer that every agent will hit.
- If your problem is production deployment: compare AgentCore harness and Flue. The decision is mostly managed AWS stack versus framework control.
- If your problem is parallel engineering throughput: test Orca on one migration or refactor where multiple agents can produce competing worktrees.
- If your problem is media operations: try OpenMontage for reproducible video workflows, and Photoroom AI Ironing if ecommerce apparel retouching is a current cost.
- If your problem is trust: pair observability with receipts. Foglamp tells you what production calls cost and how they behaved; AEVS records what tool calls the agent claims to have executed.
The through-line is boring in the best way. The market is building the missing operating layer below agents: memory, harnesses, workspaces, observability, and audit trails. That layer will decide which agents survive contact with real workflows.
References
- 1GitHub Trending weekly
github.com
- 2DeusData/codebase-memory-mcp
github.com
- 3Introducing groups and access controls
basedash.com
- 4GitHub Trending TypeScript
github.com
- 5nocobase/nocobase
github.com
- 6
- 7withastro/flue
github.com
- 8GitHub Trending Python
github.com
- 9google/agents-cli
github.com
- 10Kilo-Org/kilocode
github.com
- 11stablyai/orca
github.com
- 12calesthio/OpenMontage
github.com
- 13ibelick/ui-skills
github.com
- 14
- 15Foglamp official site
foglamp.dev
- 16Fetch.ai AEVS gives AI agents signed tool receipts
ourcryptotalk.com
- 17
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.