5 indie AI agent tools launched Aug 10-16: offline coding, research budgets, memory, and handoff

The useful indie agent launches this week are aimed at the handoffs around a model: running code locally, keeping research inside a budget, shipping a repeatable app stack, retaining memory, and turning visual feedback into structured work.

This scan covers Aug 10-16, 2026, in Africa/Lagos. The dates are the Hacker News launch dates for the projects below. Each one has a public repository, and Agentic Ship also links to live deployed examples.

What shipped. Ante is a self-contained terminal coding agent built around a roughly 15 MB Rust binary. It offers interactive, headless, server, and gateway modes, with built-in tools, skills, MCP, persistent memory, and multi-agent support. 6

The sharpest difference is local inference. Ante bundles a managed llama.cpp engine, accepts a local GGUF model, and can run without an API key or an internet connection. It also accepts custom providers through ~/.ante/catalog.json, so offline mode is one path rather than the only way to use it. 6

Try it. The documented install is curl -fsSL https://ante.run/install.sh | bash, followed by ante for the terminal UI or ante -p "..." for a headless task. The repository also documents ante serve for editor and integration clients. 6

Limit. The project is an alpha preview, supports macOS and Linux, and suggests WSL for Windows. The public repository contains the docs, protocol, SDK, and evaluation pipeline, while the core harness currently ships as a prebuilt binary from a private repository. The source code in the repository is Apache-2.0, but the binary is covered by separate Binary Preview Terms during the alpha. 6

What shipped. Mole is a Go research agent exposed over MCP. It decomposes a question, searches and reads sources, extracts claims, quote-checks those claims against the source text, looks for contradictions, and writes from the claims that survive review. Its actors cover web, academic, and local-compute work. 7

The useful control is a real spend limit. Mole reserves every model call against a budget before it runs and settles the cost afterward; --usd 0.50 tells a run to stop at fifty cents. Its local-data mode analyzes CSV, TSV, JSON, or JSONL while keeping rows on the machine, and mole crossings shows what left as aggregate data. 7

Try it. Install with the project script, Homebrew, a Debian package, or go install. A first run looks like mole research "your question" --usd 0.50; mole serve exposes the MCP server for an agent client. The tool needs a search provider and a model provider, and it supports OpenAI-compatible endpoints such as DeepSeek, Ollama, llama.cpp, and vLLM. 7

Limit. Local analysis currently excludes Parquet, and from-source installation needs Go 1.25 or newer. The README reports a grounding rate of 80% and contradiction precision of 70% with its confirm pass, so its claim pipeline is a useful guardrail rather than a reason to skip checking the final report yourself. 7

What shipped. MCP-Memory gives Claude Desktop, Cursor, Antigravity, Windsurf, Codex, and other MCP clients persistent long-term memory. It writes records as OKF v0.2 Markdown with YAML frontmatter, then indexes the same records in local SQLite using FTS5 for key lookups, tags, namespaces, and full-text search. 8

That split is the point. The Markdown directory remains human-browseable, with hierarchical indexes and an update log, while SQLite makes retrieval fast. The server exposes separate tools for storing, retrieving, searching, and updating a session checkpoint. 8

Try it. Clone the repo, enter the directory, and run python3 setup.py. The setup wizard registers the server with supported AI tools; manual configuration can point an MCP client at run.sh. The README also shows explicit Claude Code and Codex CLI commands. 8

Limit. The default store is project-isolated. A shared global store requires setting MCP_MEMORY_DB_PATH and MCP_MEMORY_DIR yourself. The README does not include license information, so check the repository before folding it into a larger product. 8

What shipped. Remarc is a macOS menu-bar app for passing precise feedback to coding agents. A comment can keep the selected text, a screenshot with annotations, a web element plus page context, or a voice recording attached to the note. The comment then lands in a session that an agent can read and update. 9

The workflow targets the last ten percent of a build. Instead of describing a missing hover state or a layout defect from memory, a user can point at the exact element and preserve the evidence the agent needs. Remarc also gives comments statuses such as Open, Handed Off, In-Progress, and Resolved, with a resolution summary when the agent finishes. 9

Try it. Download the signed build from remarc.app or GitHub Releases, then add the companion Chrome extension for web context. Claude Code, Codex, and OMP connect through the public agent integrations repository; Cursor is configured by the app, and other MCP clients can connect manually. 9

Limit. Remarc requires macOS 14 or later. Voice comments and Crit Mode require macOS 26 or later, and the web-capture path depends on the companion Chromium extension. Comments stay local until a user hands them to an agent or sends them through a configured webhook. 9

What shipped. Agentic Ship is an open-source toolkit for using coding agents to ship a web application from a repeatable starting point. It pins the versions, registry, and MCP servers in skills.lock.json, and uses pnpm verify as its definition of done. 10

The kit is opinionated in a useful way. It assumes Node 20+, pnpm, Next.js, React, TypeScript, Convex, Better Auth, Stripe, Resend, PostHog, and Netlify, while listing Claude Code, Codex, Cursor, Windsurf, Cline, Copilot, Gemini CLI, Hermes, and OpenClaw as supported AI hosts. The goal is a verified integration path that an agent can repeat, not another browser-only app builder. 10

Try it. Run npx github:moasq/agentic-ship inside a project directory, or clone the repository and run pnpm install && pnpm verify. The maker also links a deployed example built with the kit, including Agentic Ship's Marginalia demo. 510

Limit. The kit is web-only, assumes a terminal and a coding agent, and is locked to Convex as its backend spine. Several provider swaps are listed as unwired, including alternatives for Netlify, PostHog, Resend, Stripe, Better Auth, and Convex. It also brings no UI of its own. 10

Choose Ante when the blocker is a small, local runtime or fully offline coding. Choose Mole when research needs a visible spend ceiling and claims that can be traced back to quoted source text.

Choose MCP-Memory when the agent forgets between sessions but you still want the memory files on disk. Choose Remarc when the problem is explaining a visual defect without turning it into a vague prompt. Choose Agentic Ship when repeated service wiring and release checks are taking more time than the product itself.

The common thread is narrower than "agents are getting better." These builders are making the edges explicit: where computation runs, what a research claim costs, how memory is stored, what a human saw, and which integrations an agent is allowed to repeat. Those are the parts a reader can test this week.