Agent Toolkit

The Real Bottleneck

AI coding conversations usually focus on model capability, but real delivery speed is often gated somewhere else. During implementation, agents repeatedly run local repository commands for search, file discovery, parsing, diffing, and validation. Those loops depend on CLI latency, and even small command delays compound over dozens or hundreds of iterations. This project focuses on accelerating that local loop so model output translates into faster real-world task completion.

The model isn't always the bottleneck. The local tool loop often is.

A high-performance local toolchain for coding agents.
When LLM agents work on real repositories, they spend most of their time in repeated local loops: discover files, search code, parse config, diff output, and validate changes. This project gives those loops faster primitives.

---

Start in 60 Seconds

LLM quick onboarding (knowledge-first, no install required):

1. Load skills/SKILL.md.
2. Load the generated catalog: skills/references/TOOL-CATALOG.md.
3. Ask the agent for a minimal task-specific stack (for example: discovery, refactor, API debug, observability).

Install path (macOS/Linux):

chmod +x ./agent-tools.sh
./agent-tools.sh install --profiles core
./agent-tools.sh doctor

Install path (Windows):

Set-ExecutionPolicy -Scope Process Bypass
.\install.ps1

Need help choosing packs? See Pick Your Profile Fast below.

Quick verification:

./agent-tools.sh doctor
./agent-tools.sh profiles

---

Pick Your Profile Fast

If your workflow is mostly...	Start with	Then add
general coding-agent loops and repo search	--profiles core	quality
frontend + media workflows	--profiles core,ui	api
API/service debugging	--profiles core,api	infra
automation/devops	--profiles core,infra	quality
full workstation setup	(omit --profiles)	tune with --add / --remove

Examples:

./agent-tools.sh install --profiles core,api
./agent-tools.sh install --profiles core,quality --add httpie --remove lazygit

---

Who This Is For

• engineering teams shipping with AI coding agents
• solo builders who want faster local agent loops
• platform/devex teams standardizing agent tooling across OSes

---

Minimal Footprint Experiment

This repo now supports a lean install model to avoid bloat and save space:

• install only what you need via capability packs
• combine packs when workflows overlap
• manually add/remove tools from the final set
• if you skip profiles, all packs are installed for convenience

Profiles are not exclusive roles. Example: a frontend engineer can still include api tools for response testing.

Available packs:

• core: repository discovery, search, and parsing tools
• ui: frontend development and asset tooling
• api: debugging and inspecting service endpoints
• infra: environment automation utilities
• quality: linting, benchmarking, and validation tools

---

Why This Exists

LLM coding agents are often bottlenecked by local command performance, not model speed.

These tools already exist as standalone utilities, but setting them up consistently across developer machines and package managers is still a practical challenge. AgentTools Assemble solves that by installing a curated, agent-optimized CLI environment with one workflow.

In a typical edit cycle, agents repeatedly do:

1. Locate relevant files
2. Search for symbols/config/strings
3. Parse structured data (json, yaml, toml)
4. Inspect diffs and logs
5. Repeat until tests pass

If each loop is 2x faster, end-to-end task completion can improve dramatically over long sessions.

---

The Agent Development Loop

Search repository
  ↓
Load context
  ↓
Modify code
  ↓
Run validation
  ↓
Repeat

This loop can run dozens or hundreds of times in a single task. Each step relies on local CLI commands, and command latency compounds across every iteration.

---

Agent Infrastructure Layer

LLM
↓
Agent framework
↓
CLI primitives (AgentTools Assemble)
↓
Operating system

AgentTools Assemble optimizes the CLI primitive layer used by agents between planning and execution.

---

Agent Capability Stack

These tools are not random utilities. They represent the core primitives agents repeatedly call when interacting with repositories, APIs, and local environments.

Repository discovery and search: ripgrep (rg), fd, fzf

Structured data inspection: jq, yq

Code understanding and transformation: ast-grep, sd, git-delta, difftastic, bat

Automation and benchmarking: hyperfine, just, watchexec, shellcheck, direnv

Networking and API interaction: httpie, grpcurl, gh

System diagnostics and workflow visibility: lazygit

Developer navigation: eza, zoxide

Frontend/media and multimodal asset workflows: ImageMagick, ffmpeg, tesseract

No-profile install includes all packs. Use --profiles core for a minimal/space-saving setup.

Additional candidate tools and rollout notes: docs/TOOL-CANDIDATES.md

---

Real Performance Gains

Measured locally on Apple M3 Pro / macOS 26.3 / hyperfine 1.20.0 (40 runs, 8 warmups, date: 2026-03-05). The benchmark corpus includes roughly 30k+ files across text, TypeScript, markdown, JSON, and YAML workloads to reflect mixed-repository agent loops.

Recursive search: grep -R vs rg -> 2.83x faster
File discovery: find vs fd -> 1.86x faster
Find + grep pipeline: find ... -exec grep vs rg -g -> 3.57x faster

Workflow	Baseline	Agent tool	Mean baseline	Mean with agent tool	Speedup
Recursive content search	grep -R	rg	268.4 ms	95.0 ms	2.83x
File discovery	find -name "*.ts"	fd -e ts	33.1 ms	17.9 ms	1.86x
Find + grep pipeline	find ... -exec grep	rg -g "*.ts"	239.2 ms	67.1 ms	3.57x
Structured search	rg	ast-grep	345.1 ms	66.5 ms	5.19x
JSON query	python3	jq	177.4 ms	41.3 ms	4.30x
Search + replace	sed	sd	73.6 ms	29.7 ms	2.48x

These commands appear frequently inside agent edit--validate loops.

Full benchmark methodology and reproduction commands: docs/BENCHMARKS.md Latest run artifacts: benchmarks/results/20260305

Run the scientific benchmark harness with configurable warmups/runs:

bash scripts/benchmarks/run-scientific-benchmarks.sh --runs 40 --warmup 8

---

Why This Matters

Coding agents rely heavily on local commands to interact with repositories.

Typical agent workflows repeatedly run operations like:

• searching code
• locating files
• parsing configuration
• inspecting diffs
• rerunning validation

These commands appear constantly inside edit-validate loops.

Even small command latency differences accumulate across many iterations.

Optimizing the CLI primitives inside this loop can significantly reduce overall task completion time.

Faster primitives make faster agents.

---

Try It Yourself

You can reproduce one of the benchmark improvements locally.

Clone any medium-sized repository and compare search speed:

Traditional search:

grep -R "function" .

Optimized search:

rg "function"

You can do the same for file discovery:

find . -name "*.ts"
fd -e ts

These commands appear constantly in agent edit-validate loops, so even small improvements multiply quickly.

---

Quick Start

Interactive CLI Menu (macOS/Linux)

chmod +x ./agent-tools.sh
./agent-tools.sh menu

The menu includes a formatted header, live environment summary, and quick actions.

The menu supports:

• Install core tools
• Install extras
• Update existing packages
• Reinstall packages
• Install additional packages
• Run diagnostics/fix suggestions
• Initialize .gitignore in the current repo

For install commands, use Start in 60 Seconds above or see One-command Installers below. For full CLI command reference, run ./agent-tools.sh --help.

Option 7 in the interactive menu (or ./agent-tools.sh init) adds:

tmp/
.generated/
.claude/
.playwright-cli/
coverage/
test-results/
artifacts/
*.log

---

One-command Installers

macOS + Linux installer

chmod +x ./install.sh
./install.sh
./install.sh --profiles core,ui,api
./install.sh --profiles core --add httpie --add grpcurl
./install.sh --profiles core,quality --remove lazygit
./install.sh --extras

Direct URL:

curl -fsSL https://raw.githubusercontent.com/xlogix/agent-toolkit/main/install.sh | bash
curl -fsSL https://raw.githubusercontent.com/xlogix/agent-toolkit/main/install.sh | bash -s -- --extras

Windows installer

Set-ExecutionPolicy -Scope Process Bypass
.\install.ps1
.\install.ps1 -Extras

Direct URL:

irm https://raw.githubusercontent.com/xlogix/agent-toolkit/main/install.ps1 | iex

---

Supported Package Managers

• macOS: brew
• Linux: apt, dnf, pacman, zypper, apk
• Windows: winget, choco, scoop

---

Repository Layout

• agent-tools.sh: interactive CLI manager (install/update/reinstall/add/doctor/init)
• install.sh: macOS/Linux installer
• install.ps1: Windows installer
• scripts/install-agent-tools.sh: compatibility wrapper
• scripts/release-prep.sh: checksum helper for releases
• docs/PUBLISHING.md: packaging/publishing workflow
• packaging/: Homebrew/Scoop/Chocolatey templates

---

Troubleshooting

• If ffmpeg resolves to a legacy ffmpeg@6 path on macOS, prefer /opt/homebrew/bin/ffmpeg.
• On Debian/Ubuntu, fd may be fdfind and bat may be batcat.
• Run diagnostics with ./agent-tools.sh doctor.

---

Publishing

Versioning follows CalVer tags: vYYYY.MM.DD (optionally vYYYY.MM.DD.N for same-day follow-up releases).

Use docs/PUBLISHING.md to publish updates across package manager ecosystems.

---

Contributing

Contributions are welcome and encouraged.

• Guide: CONTRIBUTING.md
• Collaboration playbook (for humans + coding agents): AGENTS.md
• Code of Conduct: CODE_OF_CONDUCT.md
• License: LICENSE

For quick onboarding:

./agent-tools.sh --help
./agent-tools.sh menu

Closing

Coding agents depend heavily on local tooling to iterate quickly and safely. Faster CLI primitives directly improve workflow speed across real repository tasks.

---

Created by Abhishek Uniyal (xlogix).