Five indie AI agent builds from Aug 10–17: from research budgets to cross-device control (2026)

Most of this week's useful agent work sits beside the model. Builders are adding budgets, source checks, web extraction, traces, remote control, and review loops—the pieces that decide whether an agent can survive outside a demo. 1 2 3 4 5

Coverage: Aug 10–17, 2026, through 08:45 PKT on Aug 17.

Five open-source or indie-style projects stood out this week. They are less about launching another model and more about giving agents a boundary: a spend limit, a reliable tool layer, an audit trail, a remote cockpit, or a workflow that can check its own work.

Build	What shipped	Best reason to open it	Main prerequisite or caveat
Mole	A terminal research agent with enforced budgets, checked quotes, and a local-data boundary. 6	You need research that shows its evidence and stops spending.	Configure a search provider and a model provider; local models are supported. 1
PyScrappy	A Python scraping toolkit plus an MCP server, with adaptive selectors and structured output. 7	You want an agent to fetch changing websites through a reusable tool layer.	MCP needs Python 3.10+; browser, proxy, and stealth features are optional extras. 2
Grafana agent observability for Hermes	An unofficial plugin that records LLM calls and tool executions as OpenTelemetry traces and metrics. 8	You are debugging an agent whose output alone is too late to inspect.	Full conversation content leaves the machine by default; set metadata-only capture if that is unacceptable. 3
Zeron	A native control plane for running coding-agent sessions across machines and steering them from another device. 9	You keep long-running Claude Code, Codex, or similar sessions on another machine.	The site presents a desktop download and lists version 0.2.3; this is an early product, not a mature remote-operations guarantee. 4
Agentic Data Scientist	An adaptive multi-agent workflow that plans, executes, validates, and reflects on data-science tasks. 10	You want a data task to produce a plan and review loop, not one opaque coding pass.	Requires Claude Code plus OpenRouter and Anthropic API keys; public materials do not state team size. 5

Mole: make research spend and evidence visible

Mole appeared on Hacker News on Aug 14 PKT as a free, open-source deep-research agent for the terminal. Its author describes three promises: enforce a budget, attach a source to every claim, and keep local data on the user's machine. 6

The repository makes the mechanism unusually concrete. Mole decomposes a question, searches and reads sources, extracts claims, checks quotes against the source text, looks for contradictions, and writes a cited answer. Each model call is reserved against a ledger before it runs; --usd 0.50 means the run stops at fifty cents, and the project reports 0% measured overshoot on its test corpus. 1

It also has an MCP mode, so a coding agent can hand Mole a question and collect the answer, or use Mole as a toolkit while doing the reasoning itself. For local CSV or folder analysis, the repository says only aggregates such as counts, means, tests, and buckets of at least five records are allowed back from the local-data path. 1

Why it matters: a research agent usually fails in ways that are hard to price or audit. Mole makes both boundaries explicit. The trade-off is setup: you need a search provider and a model provider, with keys stored in Mole's local configuration. It can use an OpenAI-compatible endpoint, including a local model, but a local model still consumes tokens even when the money cost is zero. 1

Try or read: install it from the official repository, then start with a small question and a hard dollar or token limit. The interesting test is whether the citations and budget ledger change how much you trust the answer.

PyScrappy: give agents a tougher web boundary

PyScrappy was posted to Hacker News on Aug 16 PKT as a self-healing scraping toolkit with an MCP server. 7

The project accepts a URL and can return structured text, links, images, tables, and metadata. Its selector layer supports CSS and XPath traversal, while its adaptive selectors remember an element and try to relocate it by similarity when a site's markup changes. The repository also lists optional Playwright rendering, concurrent scraping, retry and rate limiting, proxy support, and 24 built-in scrapers. 2

The agent-facing part is the useful change. Install the MCP extra and an agent such as Claude or Cursor can call PyScrappy's scrapers as tools, receiving Markdown or JSON instead of raw page fragments. A built-in local-model path lets an Ollama model call the same tools directly, provided the model supports tool calling. 2

Why it matters: a web agent is only as dependable as the extraction layer below it. A selector that can adapt to modest markup changes may remove one common source of silent failure. That is a capability to test, not a guarantee: similarity-based recovery can still pick the wrong element, and blocked or JavaScript-heavy sites may need the optional browser or proxy path.

Try or read: the smallest path is pip install pyscrappy, followed by the MCP setup in the repository. For MCP, the project requires Python 3.10 or newer; the core library still works on Python 3.9. 2

Grafana agent observability: inspect the calls, not just the answer

An independent Show HN post on Aug 16 PKT introduced an unofficial Grafana Agent Observability plugin for Hermes Agent. 8

The plugin records LLM calls and tool executions as generations and emits OpenTelemetry traces and metrics. The repository's setup path is a Python package plus a direct edit to Hermes's ~/.hermes/config.yaml; it sends the resulting data to Grafana Cloud's Agent Observability surface. The author says the free Grafana Cloud tier is enough to start. 3

The privacy setting deserves more attention than the dashboard. With no capture mode configured, the plugin records full content: system prompt, tool definitions, prompts, assistant replies, tool arguments, and tool results. Setting AGENTO11Y_CONTENT_CAPTURE_MODE=metadata_only keeps the call records while excluding that content from transmission. One-shot hermes -z also disables logging for the whole run. 3

Why it matters: an agent can produce a plausible final answer after taking a bad tool path, burning tokens, or leaking more context than intended. Traces let a builder inspect the sequence that produced the result. The cost is operational and privacy-related: you need Grafana credentials, and the default capture mode requires an explicit decision before using it with sensitive work.

Try or read: start with the plugin README. Treat the first run as a privacy configuration exercise: choose metadata-only or full capture before asking Hermes to touch real data.

Zeron: take the session with you

On Aug 14 PKT, independent builder Wing Lee posted Zeron as a native, cross-device control plane for coding agents. 9

Zeron's site says it runs Claude Code, Codex, Grok, Hermes, and Pi on the user's machines, then brings their sessions into one native window. It lists live branch diffs, commit history for each workspace, and a device model in which a session started on a MacBook can be watched or steered from a phone or tablet. The site labels the product open source and MIT licensed, and shows version 0.2.3. 4

Why it matters: the remote-control problem is less glamorous than agent generation, but it is practical. A long-running local session becomes useful away from the keyboard if the builder can inspect a diff and approve the next step without reopening the whole workflow. Zeron's value is therefore tied to session continuity and device reachability, not to a new model.

Try or read: use the Zeron site to inspect the download and the supported agent list. Before trusting it with an important repository, test how it handles stale sessions, disconnection, and a rejected diff.

Cargando tarjeta de contenido…

Agentic Data Scientist: put review loops around data work

A launch post on Aug 13 PKT described Agentic Data Scientist as an open-source framework for adaptive multi-agent workflows that automate planning, execution, and validation of data-science tasks. 10

The repository describes a workflow built on Google's Agent Development Kit and the Claude Agent SDK. It separates planning from execution, validates progress against success criteria, and adapts the plan as new findings appear. It also lists MCP integration, file handling, extensible prompts and workflows, and a simple mode for direct coding when a full orchestration loop is unnecessary. 5

The prerequisite list is heavier than the pitch. The documented setup requires the Claude Code CLI, an OpenRouter key for planning and review agents, and an Anthropic key for the coding agent. The project can be installed with uv tool install agentic-data-scientist or run with uvx; network access is enabled by default but can be disabled with DISABLE_NETWORK_ACCESS=true. 5

Why it matters: separating planning, coding, review, and criteria checking gives a data task more places to fail visibly. It also makes the cost and latency harder to ignore: the full mode runs several roles and providers, while simple mode exists specifically to avoid that overhead. Public materials do not state the team's size, so this is best treated as an open-source project to inspect rather than evidence of a large production system.

Try or read: the official repository has both --mode orchestrated and --mode simple examples. Start with a small dataset in simple mode, then compare the extra review loop against the additional API and setup cost.

Cargando tarjeta de contenido…

The useful pattern: make the boundary testable

These five projects point in the same direction, although they do not prove a market-wide shift. Mole caps spend and checks evidence; PyScrappy turns web access into a reusable tool layer; Grafana exposes the calls and tool runs; Zeron extends local sessions across devices; Agentic Data Scientist separates planning from execution and review. 1 2 3 4 5

For a quick starting point: