Indie agent builds, Aug 19–24: shared brains, persistent intent, and a Lightning toll booth (2026)

The past seven days produced a useful split in indie agent work: builders are giving agents shared context, clearer intent, and tighter boundaries around access.

Keep context with the work

OzBrain — one shared brain across agents

OzBrain is a hosted, structured knowledge base that Claude, ChatGPT, Cursor, and coding agents can read and write through connectors. The product routes knowledge into linked articles, surfaces conflicts before a write is promoted, records which agent changed an article, and splits oversized articles into smaller pieces. 1

The builder's reason for making it is familiar: passing Markdown files between Claude Code, Cursor, Codex, and other agents leaves several copies in play. OzBrain keeps one current source, with a staged-write flow and an audit trail. The maintenance loop is still described as alpha, so the builder asks early users to try it before putting customer data in it. 2

Try it: start with the free plan, then connect it through the Claude setup or the MCP endpoint shown in the product guide. 1

Flocker.md — give each agent a role and a live page

Flocker.md turns agent profiles into working surfaces. Each profile has a role, context documents, a live feed, and publishing controls; an orchestrator can start specialist agents and follow their updates from a dashboard. The product lists Claude Code, Codex, OpenCode, OpenClaw, Hermes, and MCP among its supported connections. 3

The interesting shift is from a collection of prompts to a team with visible roles. A developer agent can post an update, a documentation agent can review it, and an orchestrator can read both feeds before assigning the next task. Flocker.md's free tier includes two agent profiles, live pages, custom context documents, and 20 posts per day. 3

The project appeared on Show HN on August 19, 2026. 4

Try it: create an agent profile, or follow the connection guide to publish a first post from Claude Code or Codex.

Heimdall — memory that checks the file before it trusts the hit

Heimdall is a local memory layer for coding agents that indexes work across repositories instead of keeping memory inside one project. Its search combines lexical matches, local embeddings, and graph walks, then rechecks every result against the current filesystem. Results can carry STRONG, WEAK, REBUILT, STALE, or REMOVED verdicts. 5

That recheck is the useful part for a builder with a large home directory. A remembered path can move, disappear, or match only by meaning. Heimdall records those cases instead of handing an agent a plausible dead path. The README says indexing uses tree-sitter parsing and local embeddings on the CPU, with no LLM call for maintenance. 5

The current release is v0.2.0. The package wires into Pi, Claude Code, Codex, Cursor, and Windsurf; the README marks Linux daemon packaging and MCP mode as future work. 5 The builder posted Heimdall to Show HN on August 22, 2026. 6

Try it:

npm i -g @arihantdeva/heimdall
heimdall init --harness claude-code

Ranked search also needs the Graft backend, as the README explains. 5

Make the agent follow a clearer loop

Huzzah — keep the intent beside the generated code

Huzzah is an experimental editor built around persistent pseudocode. You write a compact .hz file, save it, and the editor regenerates source code from the changed pseudocode. The pseudocode stays beside the generated code, so the human's intent remains part of the project instead of disappearing into a chat transcript. 7

The author's FizzBuzz example makes the interaction concrete: changing loop 100 to loop n changes the declaration, and Huzzah uses the diff as the next model prompt. The approach is aimed at new codebases and remains a proof of concept; cross-file dependencies, domain-specific work, and large existing projects may need a different interface. 7

The Show HN post went up on August 21, 2026. 8

Try it: start with the source code and setup instructions. The author also linked a short demonstration video. 8

Proliferate — run several native harnesses in one workspace

Proliferate is an open-source, self-hostable AI IDE from the Proliferate team. Its desktop app runs Claude Code, Codex, OpenCode, Cursor, Grok, and other coding agents in parallel. Each task gets an isolated Git worktree, branch, terminal, conversation, and review state. 9

The control surface extends beyond parallel tabs. A parent agent can delegate to subagents, and reusable workflows can pass documents between agent steps and human approval gates. The project also keeps MCPs, skills, Computer Use, Browser Use, and custom tools in one shared configuration. 9

The project is AGPL-3.0 licensed and has a self-hosting path through Docker, AWS, GCP, Azure, Kubernetes, or an air-gapped deployment. The founder introduced it on Show HN on August 21, 2026, and included a two-minute demo of the team using Proliferate to build Proliferate. 9 10

Try it: download the macOS app, read the deployment docs, or run the project from source with the README's make install and make dev-local commands. 9

Put a boundary around agent access

Argentic — an HTTP 402 toll booth for scraping agents

Argentic puts a machine-to-machine payment step in front of web access. An agent requests a URL, receives an HTTP 402 response with a Lightning invoice for 10 satoshis, pays the invoice, and receives a session token that remains valid for one hour. Later requests carry the token. 11

The design gives a scraping agent a payment path that uses neither an account nor a long-lived API key. The site publishes Python examples for requesting an invoice, paying it through a Lightning node, polling for the token, and forwarding later requests through the toll booth. 11

Argentic was posted to Show HN on August 21, 2026. 12

Try it: read the protocol examples and point a test request at the service before connecting it to a production crawler.

Ship the weird experiment

Desktop Vibe Fly — a fly that smells agent markers

Desktop Vibe Fly is a macOS fork of DesktopFly. The fork scans for files such as AGENTS.md, CLAUDE.md, .cursor/rules, and .kiro/steering, then treats visible windows, Finder rows, and desktop folders that lead to those files as scent sources. A project that is open on screen smells stronger than a closed icon. 13

A rendered 3D fruit fly viewed from above — *The repository's fly render, from the project's asset set. 13*

The upstream app simulates a circuit built from FlyWire connectome data, while the fork changes the sensory input that guides the fly around the desktop. The fork's README gives the install path as macOS 13+, Xcode Command Line Tools, git clone, ./build.sh, and ./DesktopFly. 13

The project appeared on Show HN on August 21, 2026. 14

Try it: clone kulikov0/desktop-vibe-fly on a Mac and run the build script. The repository also includes a showcase clip of the fly finding a vibecoded project. 13

The seven projects above give different answers to the same builder question: where should an agent's context, intent, coordination, payment, or attention live once the demo becomes a working tool? The answer changes with the job, which is why each project is worth trying on its own terms.