One subscription.
An army of Claude Code agents.

Vanara is a catalog of 213 specialist AI agents, skills, and packs for Claude Code — reviewers, planners, test guides, build fixers — each one tuned for a single job, all working at the same time.

$npx vanara install code-reviewer
Get started — freeBrowse all 213
A vanara coding at a terminal
project-planner
code-reviewer
security-auditor
test-author
[vanara]0:fleet1:catalog2:pricing$10/mo · all 213 items

[0] the fleet

Why a fleet beats a favorite

code-reviewer — focused pane

Specialists, not a generalist.

A single all-purpose assistant reviews your auth code the same way it writes your README. Vanara agents are narrow on purpose: the code-reviewer knows severity levels, the security-auditor knows OWASP, and neither one gets distracted doing the other’s job.

vanara cli — install

Recruit in one command.

Every agent installs into your existing Claude Code setup — .claude/agents/, hooks and slash commands wired automatically. No config files to hand-write, no docs to study. Type the name, get the specialist.

orchestration — bug fix

They fight in formation.

Orchestrator packs chain agents into pipelines: reproduce the bug as a failing test, patch to green, review before commit. You watch the transcript; the fleet does the work — and nothing merges without a reviewer’s sign-off.

Do one thing well, then compose. Each agent is a small, sharp tool; the pipeline is where they become a workflow.— the Unix philosophy, applied to agents

[1] the catalog

80 agents · 97 skills · 36 packs

[2] one door

Don’t pick the agent. Describe the task.

vanara-route — dispatch

It finds the right specialist.

With 213 items installed you shouldn’t have to remember which one to call. The vanara-route skill reads your installed roster, scores each agent on how well its stated job matches the task — domain, action, trigger — and hands off to the strongest match. You describe the outcome; it picks the expert.

vanara-route — no match

Missing an agent? We hear about it.

When nothing in the catalog fits, it doesn’t fake an answer — it records the gap locally, and vanara request turns your logged gaps into a note to the team. That’s how the catalog grows toward what people actually reach for.

Privacy first: gaps stay in ~/.vanara/gaps.jsonl on your machine. Nothing is sent until you send it — no telemetry, no phone-home.

any LLM

Skills work anywhere

The 97 skills are plain prompt & reference markdown — no model lock. Drop them into any capable model as working context.

Claude Code

Agents are built for Claude Code

The 80 agents use Claude Code’s agent format and tools, and install into .claude/. Multi-tool support is on the roadmap — we’ll say so when it ships, not before.

[3] learning

Agents that learn — twice.

code-reviewer — with memory

They remember your project.

Every agent keeps a per-project memory at .claude/memory/. Before a task it reads what it learned before — your conventions, the corrections you gave it, the pitfalls in this codebase — and after, it writes back what’s new. It stops making the same mistake twice on your code, and it runs entirely inside your Claude Code. Nothing leaves your machine.

forge — maker/checker loop

The catalog improves itself.

The items themselves are built and maintained by Forge, a meta-agent that runs a make → check → learn → reflect loop. A quality gate scores every item on depth and correctness; failures become durable lessons, and updates can never lower a score — average depth 100/100 across all 213. You pull the improvements with vanara update.

Two separate things, both honest: memory is per-project and lives on your machine; catalog updates are global and arrive when you run vanara update. The agent files themselves stay plain markdown you can read and own.

[4] trust

Built to run unattended.

code-reviewer — grounded

It shows its work.

Every agent runs under a protocol: ground each claim with a file:line, a command’s output, or the result of its own verification script — and run that script before reporting. If it can’t back a claim, it says so. No invented findings, no confident guesses.

refactoring-specialist — in bounds

It stays in bounds.

Before acting, an agent states what it will read and change, defaults to read-only, and stops for your confirmation before anything destructive — migrations, force-pushes, prod config. Every change it makes is logged to .claude/vanara-runs.log, so you always have the trail.

And it checks its own definition-of-done — tests green, no new secrets, the ask fully met — before claiming success. New to a repo? vanara doctor scans it and tells you which agents to install.

[5] governed

Capability changes and regressions are diffs, not surprises.

vanara scope --diff
$ vanara scope --diff

Capability scope CHANGED since the recorded baseline:

  ~ changed [agent] code-reviewer
      + tools: Bash
      ⚠ NEW capability: shell

Review these, then re-record if intended.
$ echo $?
1

An agent can’t gain shell access quietly.

Every installed agent declares its tools. Vanara classifies them — read, write, shell, network — and records a committed baseline. On the next update, vanara scope --diff compares against it and exits non-zero on any change, so a review that hands an agent shell or network access surfaces in the PR instead of slipping in. Scoping is a diffable file in your repo — no control plane, no cloud key service.

forge bisect
$ forge bisect

1 active regression(s) — a check that used
to pass no longer does:

  ✗ code-reviewer
      last passed at v1.0.0 (2026-07-01)
      regressed at v2.0.0 (2026-07-20)
      — 0/1 checks now pass

“Which part regressed?” has an exact answer.

Each verification run is snapshotted with the item’s version at that moment. When a check goes red, forge bisectattributes it to the exact good→bad transition — this item, since this version — instead of a vague “something got worse.”

A generic agent runner can’t do this: it doesn’t own the agent, its checks, or its version history. Vanara authors all three, so a regression is a bisect, not a hunt.

[6] the edge

What you get that a loose agent doesn’t.

On day one it looks like any folder of prompts. By the second week — when the model has shifted and your codebase has moved — the differences are the whole ballgame.

Other agentsVanara
Memory Starts cold every run — repeats the same mistakes. A per-project memory: each agent reads what it learned last time and gets sharper on your codebase.
Proof You hope the prompt worked. No evidence it did. Every item ships runnable checks (170/170 passing), and each agent self-checks its definition-of-done before claiming success.
Tokens One giant system prompt, reloaded every turn. Narrow agents; references/ load on demand — less standing context. Measure yours with vanara usage.
Trust Broad tool access, no audit trail. Capabilities are a committed baseline; vanara scope --diff flags any escalation in review.
Drift Silent rot — output degrades and nobody knows why. forge bisect pinpoints which item, and since which version, regressed.
Cost Bring your own API key, metered per token. Runs on the Claude subscription you already pay for — no keys, no metering.

[7] pricing

The whole army. One price.

$10 / month

Every agent, every skill, every update — no metering, no seats, no token bills. Agents run on the Claude subscription you already pay for. The free tier ships 30 items with no card — and it’s open source on GitHub, so you can read every line before running it. Subscribing syncs the full catalog. Cancel any time — installed items keep working.

vanara — unlock

[8] getting started

From zero to an agent army.

You need Node 18+ and Claude Code. Every agent runs on the Claude subscription you already pay for — no API keys, nothing metered.

free

Start free — no card

  1. Scan your project and get recommendations.

    $npx vanara doctor
  2. Install what it suggests — free items install for anyone, and they’re open source.

    $npx vanara install code-reviewer
  3. Open Claude Code — agents load from .claude/ automatically. Describe a task, or let an agent trigger itself.

  4. Missing a specialist? Tell us — it takes one line.

    $npx vanara request "a Bicep linter"
pro · $10/mo

Unlock everything

  1. Subscribe — your license key arrives by email instantly.

    Subscribe — $10/mo
  2. Unlock — validates your key and syncs all 213 items.

    $npx vanara unlock <your-key>
  3. Install anything in the catalog — packs expand to their members.

    $npx vanara install sre-reliability-pack
  4. Refresh the catalog any time; check your tier with account.

    $npx vanara sync

Cancel any time — installed items keep working, you just stop getting updates. Same key works on all your machines.

[9] questions

Frequently asked.

What is Vanara?

Vanara is a catalog of 213 specialist AI agents, skills, and packs for Claude Code — Anthropic's agentic coding tool. You install any of them with one command (npx vanara install <name>) and they run inside your own Claude Code.

Does Vanara use my Claude API key?

No. Vanara agents run on the Claude Code subscription you already pay for — there are no API keys to add and nothing is metered by us. Items install as markdown into your project's .claude/ folder and Claude Code picks them up automatically.

What's the difference between an agent, a skill, and a pack?

An agent is a specialist Claude Code subagent tuned for one job, like code review. A skill is a slash-command workflow that teaches Claude Code a process. A pack bundles several agents and skills for a whole discipline — one install deploys them all.

What are Claude Code subagents, and how does Vanara use them?

Subagents are scoped assistants Claude Code can delegate a task to, each with its own tools and instructions. Every Vanara agent is authored as a Claude Code subagent, so Claude picks the right specialist automatically — or you can use the vanara-route skill to dispatch to the best-fit one explicitly.

Can I use Vanara agents with other tools like Cursor or GPT?

The 97 skills are portable prompt and reference markdown that works as context for any capable model. The 80 agents are built specifically for Claude Code's agent format today; support for other tools is on the roadmap, and we'll say so when it ships.

Do the agents remember anything across sessions?

Yes. Each installed agent keeps a per-project memory file in .claude/memory/ — it reads what it learned before a task and writes back new lessons after, so it gets sharper on your codebase over time. Commit that folder to share the knowledge across your team.

Is Vanara open source?

Partly, on purpose. The 30-item free tier is open source (Apache-2.0) at github.com/vanara-agents/skills — read every line before you run anything — and the vanara CLI is Apache-2.0 on npm. The premium catalog is licensed content, like a paid font: that's what the subscription funds.

How much does Vanara cost?

30 items are free forever — no account, no card. Vanara Pro is $10/month and unlocks all 213 agents, skills, and packs, plus every update. Cancel anytime; anything you've installed keeps working.

How do I install Claude Code agents from Vanara?

Run npx vanara doctor to scan your project and get recommendations, then npx vanara install <name>. You need Node 18+ and Claude Code. Full walkthrough on the getting-started page.