New AI Tools Weekly #14: AI tools turn execution into a contract (2026)

The week in one sentence

From August 24 through August 31, 2026, AI tools made the contract around execution more explicit. They keep an append-only record, route around provider limits, validate a structured artifact, gate a deployment, or produce a proof that the output actually ran.

That pattern matters when you are choosing a first test. A model can produce an impressive answer in a short session. The surrounding tool determines whether a team can inspect the inputs, control access, recover a failed run, and hand the result to someone else.

This issue covers 12 tools from the current week's Product Hunt launch snapshot and GitHub Trending weekly scans. Product Hunt's August 31 snapshot recorded the launch-window entries as created at 07:01, and the product pages describe them as launching today. GitHub Trending supplied a separate momentum signal through weekly star gains. The two signals answer different questions: Product Hunt shows a launch event on its own surface; GitHub shows where developer attention accumulated during the week. 1 2

The practical question for every entry is the same: which boundary can you test safely before you give a tool more credentials, data, or budget?

Controlled execution surfaces

Maka and Revolte put control around agent work, with different owners for the boundary. Maka keeps execution records and permissions close to a local workspace. Revolte puts approvals, cost caps, and audit trails around a team's software-delivery process.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Apache Maka	GitHub Trending weekly: +1,973 stars. 1	Apache-2.0 open source. Desktop Nightly targets Apple Silicon Macs; the project uses your own cloud API, local model, or compatible gateway, so model charges follow that provider. 3	Maka is a local-first AI agent workspace with desktop, terminal, and evaluation surfaces. It records model messages, tool calls, tool results, permission decisions, and termination events as an append-only log. 3	Maka's useful boundary is the recoverable run record. Shortening the next prompt can reduce context while keeping the earlier tool output available for inspection. 3	Run a file-edit task in a disposable repository, deny one tool permission, abort the turn, and resume it. Check whether the record explains what happened without reconstructing the session from chat history.
Revolte Interactive Sessions	Product Hunt's August 31 launch snapshot lists Interactive Sessions as launching today. 4	Product Hunt lists free options. The public listing gives no self-serve tier price. 4	Interactive Sessions lets engineers drive architecture, code, tests, staging, and deployment step by step. Revolte also describes an Autopilot mode that takes a Jira ticket through planning, coding, pull request, and deployment. 4	Revolte keeps hands-on and hands-off execution in one governed workspace, with plan approval, inline diffs, cost caps, and audit trails. The comparison point is a coding agent that controls edits while a separate system owns delivery policy. 4	Give it a small Jira ticket against a staging repository. Require approval before the test and deployment steps, set a cost cap, and inspect the diff and audit trail before allowing the run to continue.

Maka suits a developer who wants local evidence of every action. Revolte suits a team that wants the approval boundary to travel with the software-delivery workflow.

Memory and owned context

OpenHuman and Claude Obsidian both turn past material into reusable context. OpenHuman coordinates agents, workflows, and connected accounts. Claude Obsidian keeps sources and claims as ordinary Markdown inside an Obsidian vault.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
OpenHuman	GitHub Trending weekly: +2,526 stars. 1	GPL-3.0 open source. Local workloads can use Ollama or your own provider key; managed services such as Exa-powered search and media generation use an OpenHuman subscription. 5	OpenHuman combines local-first personal memory, multi-agent orchestration, deep research, and trigger-based workflows. Its Memory Trees turn connected documents, emails, and chats into scored Markdown files mirrored in an editable Obsidian vault. 5	The differentiator is a checkpointed orchestrator around the memory store. Sub-agents can return root-cause reports, runs can be replayed with per-call cost accounting, and Privacy Mode can keep inference on the machine. 5	Feed it a synthetic project archive, create a scheduled workflow, restart the local service, and ask for the same decision through two interfaces. Delete one source and check whether the answer reflects the deletion.
Claude Obsidian	GitHub Trending Python weekly: +3,156 stars. 2	MIT open source. The workflow uses Python 3.11+, Obsidian for its visual layer, and Claude Code or a compatible Agent Skills host. 6	Claude Obsidian turns source material into linked, source-cited Markdown pages. It includes capture, query, retrieval, linting, rollup, and visual-mapping workflows for an owned knowledge vault. 6	A general vector store retrieves similar text. Claude Obsidian makes authority, freshness, support, contradiction, confidence, and review state visible in the claim record, with deterministic BM25 fallback when model-based retrieval is unreliable. 6	Add three synthetic documents with one duplicate claim and one contradiction. Query the vault, inspect the claim ledger, and verify that the conflicting evidence stays visible instead of being silently merged.

OpenHuman is the broader personal operating environment. Claude Obsidian is the narrower choice when a team needs readable files, source trails, and a vault that remains useful without a hosted database.

Model access and cost boundaries

FreeLLM API changes the route between an application and model providers. Codex changes the place where a coding agent runs. Both reduce dependence on a single interface, but they expose different operational costs.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
FreeLLM API	GitHub Trending weekly: +3,037 stars. 1	The router is MIT licensed and self-hosted. The repository describes a free local catalog snapshot and a premium live-catalog option priced at $19 per year or $49 lifetime; provider usage remains subject to each provider's limits. 7	FreeLLM API puts 34 providers and 635 model endpoints behind an OpenAI-compatible `/v1` endpoint. It supports routing, automatic failover, encrypted keys, usage tracking, and setup commands for coding agents. 7	The boundary is provider failure. A rate limit or server error can move a request to the next configured provider, while the application keeps one endpoint and one client contract. 7	Connect two disposable provider keys, set a short fallback chain, and send a fixed test set. Force one provider into a rate-limit path, then compare latency, model changes, response quality, and key exposure in the logs.
OpenAI Codex	GitHub Trending weekly: +5,510 stars. 1	Apache-2.0 open source. The CLI is available through binaries, npm, Homebrew, and other installers; ChatGPT plans are the recommended sign-in path, while API-key use follows separate setup and billing. 8	Codex is a terminal coding agent that also connects to editor extensions, a desktop app, and Codex Web. 8	Codex keeps the agent close to the local terminal while offering several ways to continue the same kind of work. The comparison point is a cloud-only coding surface whose files, tools, and credentials live outside the developer's machine. 8	Use a non-production repository with one failing test and one restricted directory. Run the same task from the terminal and an editor surface, then inspect which files changed and which permission decisions were preserved.

FreeLLM API is a routing experiment. Codex is an execution-surface experiment. Teams should measure the first against provider instability and the second against local permissions, reproducibility, and workflow fit.

Structured agent outputs

Archify and Scientific Agent Skills make the agent's working material more explicit. Archify validates a typed intermediate representation before it renders a diagram. Scientific Agent Skills packages domain procedures, tests, database access, and provenance requirements for research workflows.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Archify	GitHub Trending weekly: +18,103 stars. The repository also links version 2.16.0 dated August 30, 2026. 1 9	MIT open source. Install it as an agent skill through the public repository; the project supplies HTML, SVG, PNG, and WebM exports. 9	Archify turns a codebase or system description into architecture, workflow, sequence, data-flow, or lifecycle diagrams. Agents produce typed JSON intermediate data, which Archify compiles into self-contained artifacts. 9	Archify validates schema, layout, HTML/SVG, routes, and labels before delivery. A failed revision can return a machine-readable repair receipt while the last valid artifact stays available. 9	Give it a repository with one known dependency boundary. Compare the generated routes with real imports, deliberately break one relationship, and check whether validation rejects the artifact instead of inventing a connection.
Scientific Agent Skills	GitHub Trending Python weekly: +4,309 stars. 2	MIT open source. The collection is free; individual skills carry their own license fields, and K-Dense lists optional paid enterprise support. Installation works with `npx skills add` and several coding-agent hosts. 10	The collection provides 163 scientific and research skills covering areas such as bioinformatics, drug discovery, clinical research, databases, and scientific communication. Skills include examples, integration guides, reference material, and tests. 10	The unit of reuse is a tested procedure with retrieval and provenance rules, rather than a general-purpose prompt. The database-lookup skill uses deterministic API access and endpoint provenance to make a research workflow easier to audit. 10	Use a public, low-risk database question. Ask the agent to retrieve records, preserve endpoint and filter details, and produce a short evidence table. Verify the output against the source API before trying a clinical or proprietary dataset.

Archify checks the shape of an artifact. Scientific Agent Skills checks the method and evidence behind a domain task. Both turn a plausible answer into something a reviewer can inspect.

Media creation with editable control

The media pair moves generation closer to a production workflow. The GPT Image 2 library turns prompt examples into reusable structures. Fotor's Video Agent keeps motion-graphics elements editable on a timeline after generation.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
awesome-gpt-image-2	GitHub Trending weekly: +13,413 stars. 1	MIT open source. The library is free to browse and install as an agent skill; its companion site uses credits, with a repository migration file listing $5 for 300 credits. 11	The project organizes 544 reverse-engineered GPT-Image2 cases, 20-plus industrial templates, and composable fields for subjects, lighting, materials, layout, and visual details. 11	Prompt-as-Code gives an agent a reusable schema and style library instead of a pile of one-off prose prompts. The same data can feed the gallery, an agent skill, and a self-hosted generation workflow. 11	Pick one product-image brief and express it with the atomic schema. Generate three controlled variants, change one field at a time, and record whether the output responds to the field or drifts because the model ignores it.
Video Agent by Fotor	Product Hunt's August 31 launch snapshot lists Video Agent as launching today. 12	Product Hunt lists free options and gives no public price for this launch. 12	Video Agent turns ideas, scripts, or raw assets into motion graphics and video through chat. Fotor says text, numbers, logos, and charts remain editable on a multi-track timeline before rendering. 12	The editable timeline is the boundary. A late statistic or timing change can be made inside the composition, so the team can revise selected elements instead of regenerating the entire video. 12	Build a 20-second synthetic product explainer with one number, one logo placeholder, and one chart. Change the number and timing before rendering, then check whether the edits stay local and the final timeline remains understandable.

The GitHub project makes image instructions reusable. Fotor makes a generated video editable. The first test for each should measure control over revisions, not just first-pass visual quality.

Specialized research and creation

Godogen and Radar by Particle are narrower tools with different evidence loops. Godogen treats a running game or proof recording as the result to inspect. Radar turns a large transcribed podcast corpus into a searchable input for people and agents.

Tool	Current signal	Access / pricing	What it does	Differentiation	Try-it test
Godogen	GitHub Trending weekly: +1,033 stars. 1	MIT open source. The repository specifies external API keys for Gemini, xAI, and Tripo3D, plus Claude Code or Codex and local system packages; the project does not list a software price. 13	Godogen lets a user describe a game and directs Claude Code or Codex to build it for Godot 4, Bevy, or Babylon.js. The agent generates assets, runs the engine, and returns a live game or a 15-to-20-second proof recording. 13	A clean compile is not the final check. Godogen uses the running game or recorded clip as the feedback surface, so visible defects can drive another iteration. 13	Use a tiny game brief with one mechanic and one scene. Watch the live result or inspect the proof clip, list one visible defect, and run one correction before adding another mechanic.
Radar by Particle	Product Hunt's August 31 launch snapshot lists Radar as launching today. 14	Product Hunt marks Radar as payment required. The listing gives no public price. 14	Radar is a podcast search engine covering more than 130,000 actively transcribed podcasts, with about 20,000 new episodes added daily. Particle says its Podcast Intelligence API supports agent access through API and MCP. 14	Radar's boundary is a searchable, continually refreshed corpus rather than a single chat answer. The API/MCP path makes retrieval a repeatable input to a research workflow. 14	Search one narrow topic and save the episode titles, timestamps, and extracted claims. Compare the returned set with a manual search, then inspect whether an agent can preserve episode-level provenance through the API or MCP path.

Godogen is a generation loop with a visible artifact at the end. Radar is a retrieval layer whose value depends on coverage, freshness, and episode-level provenance.

What the six themes have in common

The 12 tools put a checkable object between a model call and the next action. Maka stores the run. Revolte stores approvals and delivery state. OpenHuman and Claude Obsidian store context as reusable records. FreeLLM API stores the provider route and fallback choice, while Codex keeps execution near local files. Archify stores a validated artifact, Scientific Agent Skills stores a tested method, and the media tools store reusable instructions or editable timelines. Godogen supplies a running result; Radar supplies a searchable source corpus.

The evidence supports a narrower conclusion than "agents are becoming autonomous." This week's tools are competing over the conditions that make an agent usable in a team: a record someone can inspect, an authority someone can approve, a format a machine can validate, a route someone can budget, or a result someone can rerun.

What to test first

Choose the first test by the boundary you need to learn:

You need recoverable agent work: start with Maka on a disposable repository. Deny a permission, abort a turn, and inspect the append-only record before resuming.
You need delivery governance: start with Revolte on one staging ticket. Use plan approval, an inline diff, a cost cap, and an audit trail before a deployment step.
You need owned context: compare OpenHuman and Claude Obsidian with synthetic documents. Test restart, deletion, duplicate claims, and contradictions before connecting personal or company data.
You need lower provider fragility: start with FreeLLM API using two disposable keys and a fixed prompt set. Measure fallback quality and key handling under a forced error.
You need a local coding surface: start with Codex in a restricted repository. Compare terminal and editor use while tracking file changes and permissions.
You need inspectable structure or research: use Archify on a read-only codebase or Scientific Agent Skills on a public database. Check the routes, filters, and provenance before trusting the output.
You need controlled media revision: use the GPT Image 2 library or Fotor with synthetic content. Change one prompt field or timeline element at a time and see whether the change stays local.
You need a visible creation or retrieval loop: use Godogen for one tiny game mechanic, or Radar for one narrow podcast question. Keep the proof clip or episode-level source trail with the result.

A first test should leave behind something a teammate can inspect: a run record, a denied permission, a diff, a source ledger, a fallback log, a validated diagram, an editable timeline, a proof recording, or a set of episode links. That artifact tells you whether the tool changes the work around the model, which is the part a short demo cannot measure.

Fuentes de referencia

1
GitHub Trending this week
github.com
2
GitHub Trending Python this week
github.com
3
Apache Maka
github.com
4
Revolte on Product Hunt
producthunt.com
5
OpenHuman
github.com
6
Claude Obsidian
github.com
7
FreeLLM API
github.com
8
OpenAI Codex
github.com
9
Archify
github.com
10
Scientific Agent Skills
github.com
11
awesome-gpt-image-2
github.com
12
Fotor Video Agent on Product Hunt
producthunt.com
13
Godogen
github.com
14
Radar by Particle on Product Hunt
producthunt.com