Nine agent tools that shipped this week: permissions, traces, and shared state (2026)

This week, makers moved the boundary around agents: permissions before shell access, traces before prose, shared state before parallel work, and local cost accounting before more autonomy. The nine launches below all landed on Show HN between 26 and 28 August, inside this issue's 21–28 August window. They are useful for the same reason: each turns a vague agent problem into a surface a builder can inspect, test, or run.

Control before autonomy

Talos puts a deterministic gate between the model and the shell

kurdman_007 launched Talos on 28 August. The agent routes shell requests through four stages — path floor, hardline, dangerous, and effect — then runs clean work in a sandbox. Every tool declares its effect in a manifest, and Talos separates reversible writes from actions that need an operator. /stop, /undo, and /autonomy 0 run without another model turn, so the control remains available while the model is confused or still generating. 1 2

The try path is unusually clear: the installer runs the test suites, installs into ~/talos, and starts nothing. That makes Talos worth opening when a coding agent needs shell access and the first question is "what can it touch?" The boundary is narrow by design. Talos expects a sandbox supplied by the host, works as a one-operator, one-machine setup, and the maker says it defends against mistaken models and tool-output injection rather than malicious models. 2

Test and repair the work

Understudy checks the trace instead of admiring the answer

neehao launched Understudy on 28 August. The Python framework wraps an agent, mocks its tools, describes a simulated user in YAML, and records messages, tool calls, and handoffs in a structured trace. Deterministic assertions can check required and forbidden actions; optional LLM judges handle qualities such as empathy. 3 4

Install it with pip install understudy[all], adapt an agent through an app adapter, mock the outside services, and run a scene. The useful change is where the assertion lands: trace.called("create_return") can verify the tool call even when the agent's prose sounds convincing. That makes Understudy a good first test harness for customer-service flows, code agents, and multi-step automation. The framework still needs an adapter and mocked tool behavior for each agent, so its value depends on the test boundary you define. 4

Opslane turns a user session into a repair queue

aray07 launched Opslane on 27 August. It captures errors and session recordings, ranks issues by affected users and recency, investigates a selected problem in a sandbox, runs tests, asks a second model to review the proposed fix, and opens a pull request after verification. 5 6

The action path is one Docker Compose file, with an MCP server for driving the workflow from a coding agent. Opslane belongs here because it connects an agent's observation surface to a change that a human can review, rather than stopping at an error summary. Start with a self-hosted instance and inspect the PR it produces. The workflow still places the repository, session data, sandbox, tests, and model review in the same operational loop, so teams should check what data they are willing to self-host before pointing it at production traffic. 6

Proval brings specialist review agents into your own Git host

seoes launched Proval on 27 August. The self-hosted code-review agent accepts OpenAI-compatible Chat Completions APIs, including local endpoints such as Ollama and llama.cpp, and connects to GitHub, GitLab, or Forgejo. For a pull request, Proval groups changed files into review units, sends those units to specialist sub-agents, and posts a consolidated review with severity groups and inline comments. It also logs reviews, replies, and token usage. 7 8

The fastest test is Docker Compose: persist /data, set an encryption key, connect a Git host, and point Proval at a model endpoint. This is a useful fit for a homelab or an internal network where review diffs should stay under the team's control. The README calls the project early and lists authentication, authorization, rate limiting, and lower false-positive rates among planned work. Pin the image and keep one replica when using its SQLite-backed deployment example. 8

Give agents shared state

Polign makes memory a typed store near the edge

Anup Talwalkar published the Polign memory writeup on 24 August, after the Show HN launch on 26 August. The open agent-memory demo puts validation, structured filters, and supersession rules in a typed store, alongside hybrid vector and BM25 search. The server keeps durable data in object storage, so a process or machine can restart while the memory remains. 9 10

Clone the agent memory demo, install polign-server, provide Go and an Anthropic or OpenAI key, then point the demo at a local directory. A preference or numeric comparison becomes a database operation instead of a paragraph the model must reinterpret. Talwalkar reports that the Wikipedia demo serves 12.5 million passages from S3 at about 37 MiB RSS and fits, with its embedder and web app, on a 2 GB ARM machine. The demo is open source; the underlying polign_db is closed source, which matters if you want to inspect or modify the storage engine itself. 10

Concord gives coding agents a shared work ledger

alexechoi launched Concord MCP on 27 August. The local-first MCP server lets Claude Code, Codex, Cursor, Gemini CLI, and Grok Build exchange messages, claim files before editing, preserve decisions, hand off tasks, and attach review evidence. Its five tools cover presence and claims, inspection, updates, transfers, and completion. 11 12

Run npm install -g @concord-ai/concord-mcp, then concord setup in a repository. Concord stores its local source of truth in .concord/, and linked Git worktrees share the primary checkout's database. The interesting part is the ownership protocol: a handoff remains with the sender until the recipient accepts it, while version checks prevent two agents from changing the same task state at once. Live delivery still depends on the receiving harness and session being reachable, and Concord sends coordination telemetry unless you disable it with CONCORD_TELEMETRY_DISABLED=1. 12

Watch the bill and the execution surface

Tare explains where Claude Code usage went

sachinneravath launched Tare on 27 August. The Claude Code skill reads the logs already on a developer's machine and answers questions about quota use, repeated context, MCP overhead, subagents, skills, background sessions, and rolling-window limits. Tare deduplicates repeated API-response records, attributes the cost of context to the tools that caused it, and recognizes the pattern of many short parallel sessions. 13 14

Install it with npx skills add kelviq/tare -g -y --copy --agent claude-code, start a new Claude Code session, and ask why the last limit arrived early. Tare runs locally; its scripts make no network connections, and its shareable summaries omit prompts, paths, contents, commands, and identifiers. The current boundary is Claude Code on macOS or Linux, with Python 3.9 or newer. Tare is the one to try when a quota surprise may be a background automation or a context-heavy workflow rather than the visible conversation. 14

shadok-ai makes scheduled Claude sessions visible

The makers of shadok-ai launched the cockpit on 27 August. It runs real Claude Code sessions side by side in separate Git worktrees, lets a person answer questions from a browser or Telegram, and schedules recurring prompts with a deterministic guard that runs before the model. A quiet guard leaves the agent asleep and costs zero model tokens, while a profile can forbid actions such as git commit and limit which secrets enter an agent. 15 16

Try npx shadok-ai with Node 20 or newer and a Claude subscription. The browser cockpit is aimed at work that can run while the maker is away: each agent gets its own worktree, and questions surface as clickable dialogs. The deployment boundary matters. The app drives commands on the host, binds to localhost by default, and needs an explicit password before it will listen beyond that machine; Telegram, secrets, and a Claude login add their own setup work. 16

KinoPipe gives media agents typed operations instead of a shell

nicniclapanique launched KinoPipe on 27 August. The hosted service exposes 48 typed video and audio operations over MCP and REST: trim, resize, captions, compression, scene splitting, and more. A validated request enters a worker, a single render produces the output, and the service returns a download URL or signed webhook. 17 18

Open the playground first, then connect the same recipe through MCP or call the REST API with an API key and idempotency key. The free tier starts with 100 credits, and the service says failed jobs refund credits. KinoPipe fits agents that need to edit media while the infrastructure keeps command execution out of the workflow. The trade is hosted processing: uploads are deleted after 24 hours, URL inputs are fetched for the job, and outputs use expiring links. 18

Where to start

Pick the tool from the failure you can already see. Talos sets a permission boundary; Understudy checks behavior; Opslane and Proval turn observed code problems into reviewable changes. Polign and Concord handle memory and handoffs between agents. Tare measures a Claude Code bill, shadok-ai keeps long-running sessions visible, and KinoPipe gives media work a typed execution surface. The common thread is practical: each project gives a builder a place to inspect the agent's next move before adding more autonomy.