walid@portfolio:~/lab/agent-building-repos$
cd../lab
02ideaSep 2026

Seven repos worth the disk space

Memory, a context layer, a real browser, orchestration patterns, security and science skills

Seven open-source projects that bolt onto a coding agent: persistent memory, a unified context store, a browser it actually drives, the orchestration patterns that hold it together, and skill libraries for security and science. Every name and star count re-checked — one is commonly mislabelled as official when it is not, and one name collides with a different repository entirely.

AgentsClaude CodeSkillsOpen sourceMCPclaude-mem ↗agent-browser ↗harness engineering ↗cybersecurity skills ↗agent-skills topic ↗
i
The one that matters most is not a tool

Of the seven, the harness-engineering collection is the one to read first if you are actually building rather than collecting. It is the thinking that makes the other six work together — state, context delivery, and the feedback loops around the agent. The lesson it argues for is that the harness matters more than the model, which is also why swapping models later costs you nothing once the structure is right.

The list

What each one is for

Star counts read from GitHub on 16 September 2026.

ProjectStarsWhat it gives your agent
claude-mem94,043Persistent memory across sessions — captures each session, compresses it, injects the relevant context into the next. The most-used and best-maintained option of its kind.
agent-browser42,694A real browser the agent drives itself: clicking, filling forms, running workflows like a person. Installs as a skill and works across several agents.
cybersecurity skills32,869817 structured security skills across 29 domains, mapped to real frameworks. Apache 2.0.
harness engineering4,097A curated collection of tools, patterns and guides for building the scaffolding around an agent. Reference material rather than a tool.
OpenViking—A context database unifying memory, knowledge retrieval and skills behind one store the agent reads and writes. Findable through the harness-engineering collection.
Diagram design—38 editorial diagram types for coding agents, self-contained HTML and SVG, explicitly avoiding the default diagram styles. Under the agent-skills topic.
Science skills—165 validated science skills plus 100+ scientific databases across biology, chemistry, medicine and drug discovery. Also under the agent-skills topic.
Install

The three with one-line installs

terminal9 lines
# persistent memory, as a plugin
/plugin marketplace add thedotmack/claude-mem
/plugin install claude-mem

# a browser the agent drives
npx agent-browser install

# the security skill library
git clone https://github.com/mukul975/Anthropic-Cybersecurity-Skills.git
!
One name is misleading, and one is ambiguous

The security library is called “Anthropic Cybersecurity Skills” but is an independent community project, not an official Anthropic release — its own README says so, and it is published by an individual account. Separately, two different repositories share the name awesome-harness-engineering with near-identical star counts, so searching the name alone is a coin flip; the one linked here is the walkinglabs collection. Neither is a reason to avoid them — just check what you are installing.

Before you install all seven

Three things worth knowing

✗Two of them overlap. The memory plugin and the context database both give your agent persistence, and running both means two systems fighting over one job. Start with the memory plugin — simplest to install, most battle-tested — and reach for the context database only when you need a single store that also holds knowledge retrieval and skills.✗The science and security libraries are enormous, hundreds of skills each. Loading everything makes the agent’s choices noisier, not better. Pull only the skills that match your actual work, and add more when a real task calls for one.✗These are community repositories, and several of them execute real code, drive a browser, or run security tooling. Skim the skill definition or the docs before turning any of them loose.
i
Free, but not free

All seven are free and open source. They also run on top of your existing agent usage, which makes building faster rather than cheaper — every skill invocation and every browser step is still model time you are paying for.

i
Provenance

Every repository named here was resolved and its star count read from GitHub on 16 September 2026. The three entries showing a dash are findable through the collections and topic pages linked above but were not pinned to one canonical repository, so no count is asserted for them. The security library’s skill and domain figures come from its own repository description: 817 skills across 29 domains.

← previous
Six analysts, one report
next →
A sales agent that stays up
← all experiments