v1.0Nium Code · private beta

The AI coding workspace,
for the terminal.

A CLI agent that plans, edits, runs tests and opens pull requests across your repo — routing each step to the lowest-cost model that will actually succeed, and metering every token in dollars.

Dollar-metered credits Hard spend caps Never trains on your code
~/acme-api·nium code
task #4821
➜
spent $0.39 / $1.20 capmodels · haiku → sonnetPR #482

One line to install · works on macOS · Linux · Windows

$npm install -g nium-code
// how it works

One loop, from goal to green pull request.

Nium Code turns a plain-language goal into a plan, a diff and validated evidence — with every model call accounted for in dollars.

  1. 01

    Describe the goal

    Write the task in plain English from your terminal. Nium picks up the repo you're in.

  2. 02

    Analyze & estimate

    The agent indexes what matters, scores complexity and quotes a cost range before touching code.

  3. 03

    Plan

    Get an implementation plan for approval — files, order of operations and expected diff scope.

  4. 04

    Execute

    Edits, tools and shell commands run in a sandbox. Destructive actions require confirmation.

  5. 05

    Validate

    Tests, linters and builds run. Failures are triaged and retried within your spend cap.

  6. 06

    Review & ship

    Reviewable diff, evidence and total cost — accept as a branch or open a pull request.

// what's inside

Everything a coding agent should already do.

Repo understanding, cost-aware routing, evidence-based reviews and controls a finance team can actually reason about.

Repository-aware agent

Indexes code, config, tests and docs. Retrieves only what a task needs — nothing more, nothing wasted.

Cost-aware model routing

Cheap models for edits and search. Frontier models when reasoning is hard. Every escalation is logged with a reason.

Plan → diff → tests

Every task produces a plan, a reviewable patch and a validation run. Accept as a branch or open a pull request.

Dollar-metered credits

1 credit = $0.01. Estimated cost before a task runs. Live spend while it runs. Hard caps you approve.

Secrets & destructive gates

Secret redaction before any external model call. Confirmations for destructive commands. Immutable activity trail.

Verified value per dollar

Optimized for completed tasks that pass tests — not raw request count. Retries and tool loops are all accounted for.

// model routing

The cheapest model that will actually succeed.

Nium Code classifies each step and picks a route. It escalates on ambiguity, multi-file work or failed validation — and records why.

  • Default to low-cost models for search, summarization and small edits.
  • Escalate to frontier models on hard reasoning, refactors or repeated failures.
  • Pin a specific tier when you need deterministic behavior on a task.
  • Every escalation and retry is logged with a reason and a cost.
router · task #4821healthy
  1. 01 Tier 0 · Fast

    haiku · llama 3.1 8b

    Search, summarization, small edits

  2. 02 Tier 1 · Mid

    sonnet · gpt-4.1-mini

    Multi-file edits, refactors, planning

  3. 03 Tier 2 · Frontier

    opus · gpt-5

    Ambiguous or repeated failures

  4. 04 Fallback

    auto-selected

    On provider failure, escalate gracefully

// transparent pricing

See what a month of shipping actually costs.

Pick a task profile and typical volume. Nium quotes before it runs and caps at your budget — no surprise bills.

task profile

routing tier

tasks / month

40

estimated cost

$7.20/ month

≈ $0.180 / task · 40 tasks

  • Quote shown before every task runs
  • Hard cap stops the agent — no overage
  • Full audit trail of models, tokens and retries
See plans
// trust

Built to sit next to your production code.

Security posture that lets you turn the agent loose on real repositories without holding your breath.

Never trains on your code

Your repositories, prompts and diffs are never used to train any model.

Secret redaction

Environment variables and tokens are stripped before any external model call.

Sandboxed execution

Shell commands run in an isolated workspace with confirmation gates on anything destructive.

Immutable audit trail

Every model call, tool invocation and retry is signed and exportable.

// faq

Questions engineers actually ask.

If we missed one, the docs go deeper.

Is Nium Code an IDE or a CLI?+

Nium Code is a CLI-first agent. It runs from your terminal against a local repository — no separate editor, no browser workspace. It writes diffs, runs tests and opens pull requests where your code already lives.

How is pricing different from other AI coding tools?+

Every credit is worth exactly $0.01. Tasks are quoted before they run and capped at a budget you approve. You pay for completed engineering work, not opaque 'requests'.

Which models does it use?+

Nium routes across a fast tier, a mid tier and a frontier tier depending on task complexity. You can pin a tier per task when you need deterministic behavior. Every escalation is logged with a reason.

What happens if a task fails?+

Failed tests trigger a bounded retry loop within your cap. If the cap is reached, the agent stops and hands back a plan, the partial diff and evidence — you decide whether to raise the cap or roll back.

Do you train on my code?+

No. Your repositories, prompts and diffs are never used to train any model. Secrets are redacted before external model calls.

500 free credits · no card required

Complete more engineering
per dollar.

Install the Nium CLI, connect a repo and let the agent handle the busywork — transparently priced, evidence-backed, reviewable as a diff.