agent-toolkit

Original:🇺🇸 English
Translated

Use when you need to find, run, chain, or save reusable CLI tools with tk instead of MCP servers.

2installs
Added on

NPX Install

npx skill4agent add asabale/agent-toolkit agent-toolkit

Agent Toolkit (
tk
)

Minimum turns and tokens. Use an existing tool first. When a raw command is repeatable, save it.

0. Bootstrap (if
tk
is missing)

Idempotent. Clone or update
~/.agent-toolkit
, then install or call that
tk
:
bash
git clone --depth 1 https://github.com/ASabale/agent-toolkit.git ~/.agent-toolkit || git -C ~/.agent-toolkit pull --ff-only
~/.agent-toolkit/install.sh
export PATH="$HOME/.local/bin:$PATH"
Or run
~/.agent-toolkit/tk
directly. Re-running is safe.

1. Use (
Find → Run
)

  1. Find:
    tk find "<intent>"
    (match plus a runnable command).
  2. Run that line with real
    --flag
    values:
    tk run <tool> --flag "<value>"
    .
    tk info <tool>
    for flags. On failure, run
    fix
    from
    stderr
    . If output has
    next
    , run it. Pipe JSON or
    tk run <chain>
    for multi-step.

2. Evolve (
Save → Check
)

When
tk find
misses and raw CLI solved a repeatable task:
  1. Save (
    $TK_ARG_<FLAG>
    inputs, compact JSON on
    stdout
    ):
    tk save <namespace>.<name> --summary "<1-line summary>" --safety <read-only|mutating|destructive> --file <script>
    Or:
    tk chain save <namespace>.<name> --summary "<1-line summary>" --steps '<json-or-names>'
  2. Verify:
    tk check <namespace>.<name> && tk run <namespace>.<name> --dry-run
  3. Prune:
    tk evolve
    reports unused, failures, overlap;
    tk prune
    archives drops.

Rules

  • Distill
    stdout
    . Named
    --flags
    only; no positional args; no hardcoded IDs, hosts, or paths.
  • Do not wrap one UNIX command; save tools that collapse steps, filter noise, or enforce safety.