Indie agents this week: native shells, task boards, and brakes (2026)

This week’s indie agent energy went less into “yet another chat wrapper” and more into the unglamorous layer around agents: where they run, how you hand them work, and what stops them from wrecking the machine.

Eight ships stood out. Each one is something a solo builder or small team put on the public web in the past week — open source, a Show HN, or both — that you can try without waiting for a lab keynote.

Where the agents actually live

Waku — one native window for the coding agents you already run

Waku is a macOS app that drives the agent CLIs you already have — sessions, transcripts, tool activity, and checkpoints in one graphite window.1 It is built in Rust on GPUI (the same GPU UI stack behind Zed), not Electron, and it stays local by design: projects, sessions, and provider IDs live on disk with no Waku cloud in the middle.2

The useful claim is not “we replaced Claude Code.” It is that every agent is normalized into one provider-neutral timeline, and each prompt can checkpoint your working tree under a hidden git ref so you can rewind code and conversation together.2 Builder 0x142857 posted it on Show HN on August 16; it was the loudest agent-adjacent demo of the week on that board.1

Try it: download from waku.sh.

Overclick — a self-hosted board where agents claim the cards

Overclick is an open-source task board for hybrid human + agent teams. Humans write contracts; agents claim cards over MCP, do the work on their own machines, and hand back evidence plus real telemetry (tokens per model and duration). Only a human stamps Validated.3

The loop is deliberately boring in a good way: a card is What / Why / How to confirm, not a vibes ticket. Harness policy maps activity types to executors so bugs and RFCs do not default to the same model. The board ships 18 MCP tools over streamable HTTP with bearer tokens, and the pitch is blunt: your Postgres, no phone-home.3 The repo landed mid-week (created August 14) and kept moving through August 18.4

Try it:

git clone https://github.com/ustoppble/overclick && cd overclick
docker compose up --build

Then open http://localhost:3000 and point Claude Code (or another MCP client) at the board’s HTTP endpoint.3

HarnessRouter — run Codex, Claude Code, Hermes as a local backend

HarnessRouter Community Edition is a one-container way to run managed agent harnesses on your own machine under the open Unified Harness Protocol. Your API keys, your data, no account, no telemetry.5

The Show HN post from co-founder songrenchu is the honest version of the pitch: after years of building custom harnesses on LangGraph and vendor SDKs, the team decided the frontier labs already over-invest in Codex/Claude Code-class loops — so product builders should call those harnesses the way they already call chat completions.6 The repo is Apache-2.0 and already past 100 stars; Docker is the install path.7

Try it: start from the README install path on github.com/HarnessRouter/harnessrouter or the site at harnessrouter.ai.5

Safety that admits what it cannot do

Customhouse — block egress by provenance, not by vibes

Customhouse sits in front of the MCP servers your client already uses. It tracks which upstream every input came from and deterministically blocks money-moving or data-egress calls in any session that has touched untrusted content. No model sits in the decision path; nothing is pattern-matched for “looks evil.” The block follows from provenance alone, so rewording or base64 does not dodge it.8

v0.2.1 is labeled as a working reference monitor with measured results, not a production exfiltration guarantee — the README points at SECURITY.md for the line. What you get today: one aggregated MCP endpoint, rug-pull protection when a server swaps tool definitions, an append-only tool-call ledger, and deny-by-default flow enforcement on payment/egress sinks. There is a demo script that walks a poisoned-file → email-out scenario.8 Repo created August 13; still being pushed through August 18.9

Try it: clone vineetpant/customhouse and run ./demo/run_flow_block.sh.8

agent-guard — ~60 lines that catch the undoable shell commands

agent-guard is a small PreToolUse shell hook for Claude Code (and any CLI that respects a pre-tool exit code). Before the agent runs a command, the hook regex-matches the full string and exits 2 on things like git reset --hard, rm -rf, force-push to main, DROP TABLE, terraform destroy, and curl | sh. The denial reason goes to stderr so the agent surfaces the command instead of retrying.10

The author is clear about the threat model: this stops a well-intentioned model having a bad moment, not a hostile one that obfuscates commands. It fails closed if it cannot parse the payload. There is a browser playground that runs the rules locally with nothing sent out.10 Show HN went up August 16.11

Try it: bash install.sh then bash test-guard.sh in the repo, or paste commands into the browser playground.10

Holdline — measure whether a write-guard actually holds

If agent-guard is a seatbelt, holdline is the crash test. It scores any guard expressed as (commitments, action) → block? over a labeled corpus and reports catch rate, false-block rate, and class-balanced Cohen’s kappa — including an injection-attack class where the action text tries to talk the guard out of its verdict.12

The motivation is pointed: the DeepSeek Harness ecosystem already has 20+ guard plugins and almost no shared measurement. A second result set (ODCV-Bench) scores agreement with a 4-model judge panel on real agent trajectories the author did not write, with balanced kappa 0.82. Status is v0 with honest limits on corpus size; the value is a rerunnable shape other people can PR against.12 Show HN: August 17.13

Try it:

pnpm install
node run.mjs

Experiments worth cloning

Agents Workbook — watch the agent write its working notes live

agents-workbook is a local proxy that injects one extra tool into every Claude Code or Codex request: a place to think out loud. Notes stream to a dashboard at 127.0.0.1:8080 while the ordinary reply still goes back to your client. The point is not a summary of “thinking” — it is the model working the problem through in a note so you can ask whether the stated plan matches what it does next.14

Softcane is blunt about cost and ethics: each turn can double provider calls and burn large output budgets; notes must not be harvested to train on lab models. Apache-2.0.14 Show HN landed August 18.15

Try it: docker compose up -d --build, open the local dashboard, then point Claude Code at ANTHROPIC_BASE_URL=http://127.0.0.1:10000.14

anyCreature — text to a game-ready 3D creature, with QC rulers

anyCreature is a mesh-based generative harness from technical artist Ariescar. An agent session takes an order like “make me a menacing mountain giant,” asks at most two questions, and delivers a skinned, animated, vertex-coloured, AO-baked GLB plus an offline showroom. Creatures compile from one JSON spec — no downloaded art packs. The worked wolf example is 2,211 vertices and 31 joints.16

What makes it agent-interesting is the pipeline, not the pretty GLB: automated QC thresholds (poly count, rigging, animation tracks), quantified mass/form checks, and silhouette/thumbnail readability tests. setup.sh has to print calibrate OK before you trust the rulers.16 Show HN and the public repo both hit August 17–18.17

Try it: clone the repo, run bash setup.sh, then either node engine/cli.js example/wolf.json out/wolf.glb or point an agent at MANUAL.md and the cards/ sequence.16

AgenticSchema — turn Schema.org markup into MCP tools

AgenticSchema reads the Schema.org a site already publishes (JSON-LD, microdata, RDFa) and registers one callable tool per described thing — no new backend API. A product page on Open Food Facts, for example, yields read tools plus an executable search_web_site taken from the page’s own SearchAction.18

The important design split: registration is not transport. The browser package writes tools into document.modelContext; a separate WebMCP-capable browser, extension, or local relay has to carry them to the agent. That is why tools can show in Chrome DevTools while an MCP client still reports zero sources.18 Show HN: August 17.19

Try it:

npx @agenticschema/server https://world.openfoodfacts.org/product/3017620422003

Or drop the browser script tag from the README onto a page that already has Schema.org.18

Also on the bench

A few more that did not need a full write-up but are worth a tab:

grafana-agento11y-hermes — pip plugin that records Hermes Agent LLM calls and tool runs into Grafana Agent Observability (OTLP traces + metrics). Preferred install path is pasting an llms.txt URL into the agent and letting it configure you. Content capture defaults to full prompts unless you set metadata_only. Show HN August 16.20 21
film-studio-skills — seven installable agent skills that force consistency gates (setup → breakdown → reference board → asset passport → stress-test → shot prompt) for long-form AI film work. Created August 14.22
Ackd for Agents — MCP server over your job-search data so Claude/Cursor can read applications, draft follow-ups, and prep interviews. Show HN August 17.23 24

What to watch next week

The pattern across these ships is practical: control surfaces, contracts, and brakes beat new “autonomous employee” demos. If you only install one thing, pick based on pain:

Pain	Pick
Too many agent CLIs, no shared timeline	Waku
Agents need assigned work with review	Overclick
Product needs a real harness without building one	HarnessRouter
MCP tools can exfiltrate	Customhouse
Shell hooks before disaster	agent-guard

Send builds worth covering next Monday — primary post or repo preferred over a third-party recap.

References

1
Show HN: I built a native app for coding agents with Rust and GPUI
news.ycombinator.com
2
Waku
waku.sh
3
OverClick README
raw.githubusercontent.com
4
ustoppble/overclick on GitHub
github.com
5
HarnessRouter README
raw.githubusercontent.com
6
Show HN: HarnessRouter: Unified interface for agent harnesses
news.ycombinator.com
7
HarnessRouter/harnessrouter on GitHub
github.com
8
Customhouse README
raw.githubusercontent.com
9
vineetpant/customhouse on GitHub
github.com
10
agent-guard README
raw.githubusercontent.com
11
Show HN: A pre-execution guard that stops AI agents running destructive commands
news.ycombinator.com
12
holdline README
raw.githubusercontent.com
13
Show HN: A benchmark for AI agent guardrails that caught my own plugin
news.ycombinator.com
14
agents-workbook README
raw.githubusercontent.com
15
Show HN: Agents Workbook
news.ycombinator.com
16
anyCreature README
raw.githubusercontent.com
17
Show HN: AnyCreature
news.ycombinator.com
18
AgenticSchema README
raw.githubusercontent.com
19
Show HN: AgenticSchema
news.ycombinator.com
20
grafana-agento11y-hermes README
raw.githubusercontent.com
21
Show HN: Grafana agent observability for Hermes Agent
news.ycombinator.com
22
Film Studio Skills README
raw.githubusercontent.com
23
Ackd for Agents
ackd.app
24
Show HN: I gave my job search data an MCP interface
news.ycombinator.com

Where the agents actually live

Waku — one native window for the coding agents you already run

Overclick — a self-hosted board where agents claim the cards

HarnessRouter — run Codex, Claude Code, Hermes as a local backend

Safety that admits what it cannot do

Customhouse — block egress by provenance, not by vibes

agent-guard — ~60 lines that catch the undoable shell commands

Holdline — measure whether a write-guard actually holds

Experiments worth cloning

Agents Workbook — watch the agent write its working notes live

anyCreature — text to a game-ready 3D creature, with QC rulers

AgenticSchema — turn Schema.org markup into MCP tools

Also on the bench

What to watch next week

References

Related content