Indie agent builds, Aug 25–31: local sandboxes, shared rooms, and harder boundaries

The week’s indie agent builds move the hard parts out of the chat window: local execution, shared state, repeatable tests, and boundaries around credentials and syscalls.

What shipped. HermOS is an open-source desktop IDE for multi-step coding missions on a local workstation. Its workers edit files, run sandboxed terminal processes, control browsers, and fork subagents for parallel investigation and testing. Local SQLite stores token and latency telemetry, while atomic checkpoints and visual diffs let a builder review changes before committing them. 1 HermOS appeared on Show HN on August 31, 2026. 2

Why it matters. HermOS puts autonomy and inspection in the same local surface. A builder can give an agent a longer mission while keeping the repository, usage data, diffs, and rollback points on the workstation. The project also supports MCP and local inference through Ollama, vLLM, and LM Studio. 1

Try it. Download HermOS for Windows, macOS, or Linux, or run it from source. The project says it is MIT-licensed and requires no account. 1

What shipped. Hillock is a local neuro-symbolic memory engine that combines SQLite knowledge-graph triples, Hebbian associative recall, and a 10,000-dimensional hypervector space. Its TALON pipeline extracts and gates information locally; a hard similarity threshold blocks an unanswerable query before Ollama generates a response. The README describes ingestion in about five seconds and a footprint below 1.2 GB of VRAM, with CPU-only execution available. 3 The project appeared on Show HN on August 31, 2026. 4

Why it matters. Most agent memory discussions end at retrieval. Hillock makes the next decision explicit: pass verified facts to the language model, or refuse when the local gate cannot support an answer. That makes the failure mode part of the memory design. The project’s own benchmark section calls its test set small, so the published percentages are best read as development measurements rather than a general accuracy claim. 3

Try it. Clone roandejager/Hillock, then use the repository’s launcher or manual setup. You need Python 3.10+, a local Ollama model for final response generation, and an NVIDIA GPU only when you want the faster CUDA path. 3

What shipped. Murmell is a browser canvas where multiple coding agents run on one cloud machine and share one project directory. Each canvas keeps its machine, branches, development server, and session state alive while the builder is away. File reservations use exclusive leases with a time-to-live, and a watcher surfaces writes that collide with another agent’s claim. 5 Moss'Ab introduced the project on Show HN on August 30, 2026. 6

Why it matters. Parallel agents usually create a coordination problem before they create a productivity gain. Murmell moves ownership into the workspace: an agent claims a path before editing it, another agent sees the denial and its remaining lease, and a human can watch the same terminals and previews from a second browser. The hosted machine also keeps the project alive after the laptop closes. 5

Try it. Open the short demo without committing to a project, or enter a canvas and connect the agents you already use. Murmell is a paid hosted product; the builder says it charges for the machine and collaboration rather than model tokens. 5

What shipped. Cogram Studio is a research-preview CAD and BIM workspace for humans and agents. It runs FreeCAD 1.1 headlessly with the Open CASCADE geometry kernel, exposes an MCP server for Claude Code, Codex, ChatGPT, and other clients, and supports formats including STEP, IFC, STL, DXF, and FCStd. Studio also has a built-in agent based on Pi and a browser interface for inspecting models, measurements, views, tables, and sheets. 7 The Cogram co-founders posted the project on August 31, 2026. 7

Why it matters. CAD gives an agent several representations to keep consistent: geometry, drawings, dimensions, and exported files. Studio makes those representations available through one workspace instead of asking an agent to describe a model in text. The builders say the current product works best for iterative, bounded tasks; a complex model left to run unattended can look convincing while failing closer inspection. 7

Try it. Open Cogram Studio and use the agent connection instructions with Codex or Claude Code. The Show HN post also points to a built-in-agent route with 50 free credits. 7

What shipped. Understudy is a scenario-testing framework for AI agents. A builder wraps an agent through an ADK, LangGraph, or HTTP adapter, mocks the tools that would otherwise call real services, writes multi-turn scenes in YAML, and runs simulations that record messages, tool calls, and handoffs. Deterministic checks and optional language-model judges then evaluate the structured trace. 8 The project appeared on Show HN on August 28, 2026. 9

Why it matters. A polished answer can hide a bad action sequence. Understudy gives a test a concrete target such as a required tool call or a forbidden refund, then compares that behavior across repeated simulations and versions. The same trace format also lets a builder separate generating test runs from evaluating them later. 8

Try it. Install the package with pip install "understudy[all]", then start from the repository’s example scenes and tests. The project requires Python 3.12 or newer. 8

What shipped. Grith is a Linux security supervisor for AI coding agents. It intercepts syscalls and scores file reads, shell commands, network calls, and process spawns before the kernel executes them. Scores below 3.0 are allowed, scores from 3.0 to 8.0 wait in a review queue, and scores above 8.0 are denied. The repository includes profiles for common coding agents and an audit log that stays in local SQLite by default. 10 The project appeared on Show HN on August 28, 2026. 11

Why it matters. Prompt instructions sit inside the same process that is trying to complete the task. Grith moves the decision to an OS boundary, where a malicious or confused tool call can be queued or stopped before it reaches the filesystem or network. The current release path targets Linux x86_64 and aarch64; the repository lists macOS and Windows backends for a later version. 10

Try it. On a supported Linux machine, install Grith or download a binary from the latest GitHub release, then wrap an existing agent with grith exec -- claude-code "fix the failing test". 10

What shipped. Leadcode is a macOS app for developers who work across several clients. Each persona gets its own provider configuration, GitHub account, and cloud identity. The router scrubs ambient credentials, injects only the declared binding, and launches the provider’s real CLI on a PTY. Sessions survive focus changes, while secrets stay in the macOS Keychain and the app stores local state in SQLite. 12 The project appeared on Show HN on August 28, 2026. 13

Why it matters. A wrong account can turn a harmless agent run into a billing, authorship, or deployment incident. Leadcode makes the selected identity part of the workspace rather than a shell variable that a builder has to remember before every session. The current app supports macOS 13 and newer on Apple Silicon and Intel; the project describes process isolation and lists a stronger sandbox boundary as future work. 12

Try it. Download Leadcode and create a persona for each client. The site says the first sign-in starts a 14-day Pro trial with no card, while the free tier keeps one live account per provider after the trial. 12

The practical split this week is clear: local IDEs keep execution near the repository, hosted rooms keep agents in one shared state, and the testing and security tools make their actions inspectable. Pick the project that matches the failure you are currently paying for, then try the smallest path each builder exposes.