DEV Community

Cover image for Give your agents tools they would love to use
Sumit Mishra
Sumit Mishra

Posted on Originally published at github.com

Give your agents tools they would love to use

#ai #opensource #devtools #agents

Every agent framework ships the same starter kit: a system prompt, a loop, and permission to run shell commands.

Then we wonder why it git pulls the wrong branch, burns tokens deciding whether you meant pytest or ruff, and confidently edits a file that does not exist.

The agent is not stupid. We gave it bad tools.

A model with only bash is like a brilliant engineer locked in a room with a screwdriver and no manual. It will try. It will improvise. It will eventually break something you cared about.

The fix is not a longer system prompt. The fix is a toolbox the agent actually wants to reach for.

That is what Arka is built to be.

The problem: agents improvising instead of executing

Picture what we ask agents to do every day:

  • Check repo health before a PR
  • Run the right test command for this stack
  • Look up calendar conflicts, PDFs, or web docs
  • Remember that the release branch is release/2.4 until Monday
  • Stay inside the project boundary — no drive-by git init, no hallucinated paths

Now picture what we actually give them:

What the agent needs What it usually gets
Repo map ls -R and hope
CI / lint gates Raw shell + guesswork
Structured review git diff dumped into context
Safe automation Unrestricted bash
Routing An LLM call per keystroke

The result is familiar:

  • Token burn on tasks a five-line script could handle offline
  • Context loss between IDE chat, terminal, and browser
  • Trust erosion when the agent "explores" with destructive commands
  • Fragile loops — the agent re-derives the same facts every turn

We do not need agents that talk better. We need agents with tools they would love to use — fast, deterministic, scoped, and composable.

The solution: a skill layer agents actually reach for

Arka is a local-first CLI agent with 70+ built-in skills and 120+ symbolic routing rules that resolve most requests before any LLM is called.

Install:

pipx install "arka-agent[chat]"
arka setup
arka doctor
Enter fullscreen mode Exit fullscreen mode

From there, plain English maps to real capabilities:

arka map this repo
arka health
arka review staged changes
arka ask "what changed in Python 3.13 asyncio since 3.11?"
arka goal "fix the first failing pytest in tests/"
Enter fullscreen mode Exit fullscreen mode

No hosted account. No demo instance holding your code. Config and keys live on your machine — ~/.config/arka/ on Linux, Application Support on macOS, %APPDATA%\arka\ on Windows.

Your terminal becomes the agent's workbench — not a place it wanders blindly with cd and find.

Tools agents love: fast, named, and offline-first

1. Routing that does not waste tokens

Most stacks use the LLM as the front door. Arka uses it as the fallback.

arka route "run ruff on changed files"
Enter fullscreen mode Exit fullscreen mode

Symbolic rules match weather, CI, repo map, passwords, YouTube transcripts, currency, and dozens more — zero tokens to decide what you meant. When language is ambiguous, Arka fails over across 24+ providers (Gemini, Groq, Ollama, OpenRouter, …) with keys you configure.

Sensitive work can stay local:

arka run-only-local-llm "explain this diff"
arka hybrid config local-only
Enter fullscreen mode Exit fullscreen mode

Agents love tools that are instant and predictable. Routing is the first one.

2. Skills that match how engineers actually work

Arka ships skills agents reach for constantly:

Skill family What the agent gets
Code & repo repo_map, repo_health, review, ci, pr_check, security
Data & research ask, search, pdf_rag, data_ask
Media & voice youtube, compose_video, TTS, transcription
Infra deploy, Railway, Docker, cloud helpers
Memory remember, recall, session context

These are not vague "use bash" instructions. They are named tools with clear inputs — the difference between an agent that executes and one that performs.

2b. The skills you underestimate: weather, news, and smoke tests

Some skills look personal. They are not.

weather, stock news, headlines, and quick web lookups are often the best first commands when you stand up a new agent — a fresh MCP connection in Cursor, a custom wrapper, a hosted API profile, or a skill.json plugin you just dropped in.

Why? Because they are honest integration tests:

  • No repo context required — you do not need a codebase loaded to know if routing works
  • Fast feedback — symbolic rules match in milliseconds; you see a real answer, not a model guessing
  • End-to-end proof — prompt → router → skill → output, without touching git or pytest
  • Safe — if the agent is miswired, you get a failed forecast, not a deleted branch

When I wire a new agent client, I do not start with goal "refactor the auth module". I start small:

arka route "weather in Bangalore"
arka weather tomorrow
arka stock news
arka capabilities
Enter fullscreen mode Exit fullscreen mode

If weather routes to hyperlocal_weather and returns a forecast, the plumbing works. If news hits search or your taught route and comes back with headlines, dispatch works. Only then do I trust the same path with repo health, CI, or autonomous coding goals.

Personal utilities and builder smoke tests are the same tools. That is a feature — not a footnote.

3. Guardrails so the toolbox is safe to open

Agents with great tools still need boundaries. Arka defaults to:

  • Blocks on destructive shell patterns
  • [y/N] prompts for risky installs, deletes, and downloads
  • Sanitized web content before it hits a model
  • Coding-session mode that keeps browser opens and creative skills out of the way

You are not babysitting every command. You are giving the agent tools it can use without burning the house down.

4. MCP — export the toolbox to any agent

Your IDE agent should not re-implement repo health. It should call the same skills:

arka mcp serve
Enter fullscreen mode Exit fullscreen mode

Arka exposes its skill surface over MCP (stdio / SSE) for Cursor, Claude Desktop, Copilot, and anything else that speaks the protocol. One router. Many agents. Same toolbox.

arka gemini "explain this regex"    # wrap other CLIs too
arka connector status               # shared Agent Hub context
Enter fullscreen mode Exit fullscreen mode

Give every agent in your stack the same tools they would love — not a different ad-hoc script per client.

A day in the life: an agent with the right tools

Morning — orientation without improvisation

You land in an unfamiliar repo. The agent does not guess the tree:

arka map this repo
arka health
Enter fullscreen mode Exit fullscreen mode

Structure, test entry points, lint gates — from local inspection, not invented paths.

Midday — coding with a composer, not chaos 

arka dev tui .
Enter fullscreen mode Exit fullscreen mode

The coding TUI opens with session status (branch, model, loaded skills). Plain text builds a plan; you approve with y. /test runs read-only pytest and auto-fixes once on failure. /ship runs review plus changed-only CI before commit.

The agent has verbs (/plan, /run, /test, /ship) instead of improvising shell pipelines.

Evening — memory that persists across sessions 

arka remember "release branch is release/2.4 until Monday"
arka recall release branch
Enter fullscreen mode Exit fullscreen mode

Agents love state they do not have to re-derive every turn. Arka keeps memory local by default.

Build your own tools (agents love extensibility)

Agents outgrow fixed toolkits. Arka accepts third-party skills via skill.json plugins — no fork required.

Teach routing:

arka route learn "morning standup" "stock news"
Enter fullscreen mode Exit fullscreen mode

Preview before you trust:

arka route "deploy to railway"
Enter fullscreen mode Exit fullscreen mode

The agent ecosystem gets better when builders add tools, not when they lengthen prompts.

The real insight

The next generation of agents will not win on charisma. They will win on tooling.

  • LLMs are the reasoning layer
  • Skills are the execution layer
  • Routing is the efficiency layer
  • Guardrails are the trust layer

Arka bundles all four into one CLI you already have open.

Try it in five minutes 

pipx install "arka-agent[chat]"
arka setup
arka doctor
arka route "weather in London"    # smoke test: routing + skill dispatch
arka weather tomorrow             # smoke test: end-to-end answer
arka capabilities
arka "what can you do with this repo?"
Enter fullscreen mode Exit fullscreen mode

Star the repo if it clicks — it helps other builders find an agent stack that ships with real tools:

gh repo star Sumit884-byte/arka
Enter fullscreen mode Exit fullscreen mode

What's next

Arka is actively evolving: richer coding TUI, more symbolic routes, tighter MCP ergonomics, and hosted coding profiles for Railway and beyond.

The north star stays the same: give your agents tools they would love to use — fast, local-first, composable, and under your control.

Repository: github.com/Sumit884-byte/arka

Docs: arka-agent.mintlify.site

Install: pipx install "arka-agent[chat]"

Arka — your terminal, upgraded.

Top comments (0)