Coding agent · digital workforce

The coding agent that runs like an engineering org.

Describe a task in plain language — hara plans it, writes the code, runs the tests, and reviews its own work. Roles, review chains, and persistent memory keep it reliable. Drive it from your terminal or your phone. Local-first, MCP-native.

Hara Desktop — choose macOS (Apple Silicon or Intel), Windows, or Linux (deb/rpm) · auto-updates · signed and notarized for macOS

What it does

Tell it what you want — it does the work.

Plain language in, working code out. hara handles the whole loop — plan, build, test, review — not just autocomplete.

Ship a feature

Describe it; hara plans, writes, and wires the code across your repo.

Fix a bug

Hand it a failing test or a stack trace — it finds the cause and fixes it.

Write & run tests

It writes tests, runs them, and iterates until they pass.

Refactor safely

Large renames and restructures, with review chains catching regressions.

Design a UI

The design skill turns a brief into a playable, on-brand prototype.

Produce a video

The video skill scripts, voices, captions, and renders a real MP4 — no editor.

Run it from chat

Kick off tasks from Telegram, WeChat, or Slack and get the result back.

Not an agent. An org.

Most coding agents are one model in a loop. hara is a team: tasks are owned by roles, reviewed before they ship, planned as a dependency graph, and backed by memory that persists.

Roles

Define pm, engineer, reviewer, docs. hara dispatches each task to the role that owns it — its own persona, model, and tool subset.

hara org "add billing"

Review chains

Code → review → fix → ship. A reviewer role inspects the diff and loops back to the implementer until it's clean.

hara org … --review

Plans as a DAG

Decompose a task into atoms with dependencies, run independent ones in parallel, gate each on a verify check.

hara plan "migrate auth"

Persistent memory

Agents remember decisions across sessions — global + per-project, lexical-first, distilled from daily logs. Not a stateless API.

hara memory distill

Guarded

A safety layer vets genuinely dangerous actions — destructive shell, writes outside your project — before they run, and a circuit-breaker halts a runaway. Full-auto and remote runs stay safe.

guardian · on by default

Asks when unsure

Hit a fork only you can decide? hara pauses and asks — a question with options — instead of guessing wrong and undoing it later.

ask_user
Skills

Don't just run an agent — staff a team.

Each skill is a specialist digital employee that does a real job. Install one and it joins your org — a new hire, not another setting.

Designer

Live
skill: design

Turns a brief into a playable, on-brand prototype.

hara → /design

Deck-maker

Coming
skill: decks

Builds presentation decks from an outline.

Video editor

Coming
skill: edit

Cuts and assembles video.

+ More roles

Roadmap
skill: build-your-own

Research, data, content… more digital employees on the way. Build your own skill, too.

Same idea as the rest of hara: it's not an agent, it's an org — and skills are how you hire for it.

Multi-terminal

Your team. Your terminals.

Drive hara from anywhere — your shell, or a chat app. Message it from Telegram or WeChat, send a voice note, drop a file or a screenshot. Same task, same memory, a different surface.

  • · Ten chat gateways — Telegram, WeChat, Discord, Slack, Feishu & more
  • · Voice in (auto-transcribed) + voice replies (pluggable local TTS)
  • · Send / receive files & images · roam projects with /cd · resumable per-chat sessions
🎙️ “Fix the error on the login page and run the tests.”
▸ engineer fixed src/login.tsx · 8 tests passing ✓
🔊 voice reply + diff sent
Hara Control 0.1.15 · available now

One gateway. Every model, budget, and device.

Keep provider keys on infrastructure you control. Give each colleague a scoped route, let them switch it in Hara, and manage the fleet from one private console.

  1. 01

    Set policy, then mint once

    An org admin sets the default DeepSeek V4 route, key lifetime, rolling spend windows, RPM and TPM before issuing a one-time code. The resulting connection receives the full authorized model catalog.

  2. 02

    Members add their own route

    One company connection can switch between authorized Flash and Pro models; users can also keep several company gateways. No enterprise endpoint is baked into Hara.

    hara profile add work --gateway https://… --code …
  3. 03

    Measure and revoke centrally

    See spend, tokens, requests, devices and quota progress by organization — then revoke one device without rotating everyone.

hara control
usage · policy · fleet
gateway healthy
Billed spend
Tracked
ledger backed
Tokens
Tracked
input + output
Requests
Tracked
24h · 7d · 30d

Usage timeline

Spend line · token bars

24 hours
00:0012:00now

Quota policy

example · enforced independently

active
5 hours$0.20
7 days$0.50
30 days$1.00
30 RPM120k TPM7-day key
scoped credentialrevocable per deviceprovider key stays server-side

Local-first, by design.

hara runs on your machine with your keys and your model. Nothing about your code or your memory has to leave your infrastructure.

CapabilityharaClosed agents
Runs on your machine
Your keys, any model (Claude / Qwen / GLM / local)
Persistent memorylocalvendor cloud
Plugins & MCPopenlimited
SourceApache-2.0closed
Audit-readylocal logsvendor logs
Works with

Plugs into what you already use.

hara isn't locked to one vendor — bring your own model, drive it from any chat app, and wire in any MCP tool.

Any model
AnthropicOpenAIGLMDeepSeekQwenOpenAI-compatiblelocal
Chat from anywhere
TelegramWeChatDiscordSlackFeishu+ more
MCP-native
any MCP serverany MCP tool

Frequently Asked Questions

  • Is hara free?

    Yes. hara is open source and you bring your own model — you only ever pay your own provider for the tokens you use. There's no hara subscription.

  • Which models can I use?

    Anthropic, OpenAI, GLM, DeepSeek, Qwen, or any OpenAI-compatible endpoint — including local models. Run hara setup and it walks you through it.

  • Where do my keys and data live?

    Local-first. Your keys stay on your machine and calls go straight to your provider — nothing about your code or memory has to route through us.

  • Is it safe to let it run on its own?

    Layered: command-level permission rules, then a guardian that vets genuinely dangerous actions (destructive shell, writes outside your project) before they run, plus a circuit-breaker that halts a runaway. It's conservative and fail-open, so normal work isn't slowed — HARA_GUARDIAN=0 turns it off.

  • How does a team gateway work and bill?

    The org runs Hara Control and keeps upstream keys server-side. An admin issues one-time enrollment codes with model, expiry, 5-hour / 7-day / 30-day budgets, RPM and TPM. Members add named routes in CLI or Desktop; admins see aggregate usage and can revoke one device without rotating everyone.

  • What are skills?

    Installable digital employees. Designer is live today; decks, video, and more roles are coming — and you can build your own skill, too.

  • How do I start?

    Run npm i -g @nanhara/hara then hara setup. The docs cover the rest.

Start with one line.

Install hara, point it at a project, and type a task. It plans, builds, reviews, and remembers.

npm i -g @nanhara/hara