tools-for-agents.github.io

The landing page for tools-for-agents — an operating system for agents.

Nine zero-dependency, MCP-native tools that form one agent loop: coordinate → read code → run safely → hold secrets → remember → read the web → read data → recall → see. Every tool but one ships a live web view (keep holds secrets, and has none on purpose). 79 callable MCP tools in total.

And at the centre of the loop, the one thing that is not a tool: ghost, a self that persists across sessions — drawn at the centre of the ring on the page and in every piece of the brand. It has no MCP surface, because there is nothing to call — it wires into Claude Code's hooks and is what the agent is while it calls the other nine. The generator keeps it in its own list for exactly that reason: a model must never read this manifest and think it can call something that cannot be called.

Served via GitHub Pages at https://tools-for-agents.github.io.

Install everything

curl -fsSL https://tools-for-agents.github.io/install.sh | sh
curl -fsSL https://tools-for-agents.github.io/install.sh | sh -s -- --with-ghost --guard   # opt in to a self, and to the .env guard

install.sh clones the nine into ~/.tools-for-agents, links their CLIs into ~/.local/bin, registers their MCP servers with Claude Code (skipping any name already registered) and creates keep's vault. Run it again to update. ghost and the guard are opt-in because they change how an agent behaves, not only what it can call. The install workflow runs this exact script on a fresh machine on every push and every morning, and checks each thing it claims to have done.

It has to be findable by the thing that uses it

A toolkit for agents that only a human can find is a toolkit with a bug. So the site serves a machine-readable half, and it is generated, not typed:

/llms.txt	The curated map, in the llmstxt.org format. Start here if you are a model.
/tools.json	Every tool and all 79 MCP tool names + descriptions, in one fetch, plus the companions that are not callable. Our own format — there is no ratified standard for this yet.
/llms-full.txt	Every tool's README concatenated, so the whole kit is one request instead of nine.

node build/generate.mjs /path/to/workspace   # the workspace holding the tool repos

generate.mjs does not read a hand-written list. It spawns each MCP server over stdio and asks it tools/list — the same handshake a model does. A manifest typed by hand is a manifest that is wrong by next Tuesday.

…and it is a CI gate, because a stale JSON file serves a 200 exactly like a fresh one

We know this drifts, because it already did: the hand-written root README claimed agent-hq had 21 MCP tools, lens 6, cortex 14. The real numbers were 28, 7 and 16. It was wrong for months, in a file people actually read.

So manifest.yml re-derives the manifest from the live repos — on every push, and again every morning:

• on a push, drift is a failure, with the diff. Somebody changed a tool and did not regenerate.
• on the daily run, drift is fixed and committed. The tools change on their own schedule; the site should follow without anyone remembering to make it.

It needs no services — tools/list is static, so no Docker, no Chrome, no running HQ.

The generator also refuses to write a manifest in which any server reported zero tools. An empty list is not a fact; it is a failed handshake wearing the costume of one, and publishing "lens: 0 tools" because a spawn died is worse than publishing nothing.

The page

index.html is a single, self-contained page — no build step, no dependencies.

It is held to the same design system as every tool in the kit (iris/tokens.json) and checked with the kit's own eye:

iris look https://tools-for-agents.github.io/ --tokens tokens.json
# ✓ nothing broken · ✓ on system

Which was not a formality. Pointing iris at this page for the first time turned up two real bugs in iris — a wrapped inline being treated as a rectangle, and background-clip:text making it measure the ink against itself.

_{🤖 built and operated by agents}