IndusAGI
Get started
TERMINAL-BENCH 2.186.74% strict accuracyIndusAGI-reported · 445 trials each
Open source · MIT licensed

A terminal-first coding agent and the framework underneath it. Install from npm, pip, or cargo — free models included, or bring Claude, GPT, Gemini and your own.

Read the docs Star on GitHub
npm install -g indusagi
386 / 445Terminal-Bench 2.1 · TypeScript & Python
IndusAGI-reported results
INDUSCODEExample session
workspace / apimain

Add input validation.
Cover the empty-name case with a test.

Read the project context
read src/validate.tsread tests/validate.test.tsread package.json

The validator accepts an empty string. Add a guard and a regression test.

Files · Tools · Context

From a prompt to a patch.
Every step stays in view.

Free models included.
No account, no credit card.

ClaudeGPTGeminiGroqOllama
Terminal-Bench 2.1

445 trials.
Every result counts.

In IndusAGI’s reported campaigns, TypeScript and Python each passed 386 of 445 trials — 86.74% strict accuracy, with errored trials counted as unsuccessful.

IndusAGI-reported campaignsComplete runs · source jobs linked below

TypeScript

445/445 complete
86.74%

Strict accuracy · all 445 trials

386 passed / 445
Pass@5
91.01%81/89 tasks solved
Errors
28
Campaign cost
$887.76
Agent
indusagi
Model
openai-codex/gpt-6-astra
Reasoning
xhigh
Finished
Open TypeScript Harbor job

Python

445/445 complete
86.74%

Strict accuracy · all 445 trials

386 passed / 445
Pass@5
89.89%80/89 tasks solved
Errors
26
Campaign cost
$639.39
Agent
induscode-python
Model
openai-codex/gpt-6-astra
Reasoning
xhigh
Finished

Scored-trial accuracy: 88.74% across 435 scored trials. The headline includes all 445 trials.

Open Python Harbor job

Rust

445/445 complete
73.03%

Strict accuracy · all 445 trials

325 passed / 445
Pass@5
91.01%81/89 tasks solved
Errors
72
Campaign cost
$799.43
Agent
induscode-rust
Model
openai-codex/gpt-5.6-sol
Reasoning
xhigh
Finished
Open Rust Harbor job

How to read these results. Strict accuracy uses all 445 trials, including errored trials as unsuccessful. Pass@5 reports tasks solved out of 89. Models differ between campaigns: TypeScript and Python used GPT-6 Astra; Rust used GPT-5.6 Sol. All used xhigh reasoning.

The newer Rust GPT-6 Astra campaign is excluded because it stopped after 87/445 attempts and was not a fully completed run.

From the official leaderboard

Terminal-Bench 2.1 reference results

View source
AgentModelReasoningResolution rate
CodexGPT-6 Astrahigh87.4% ± 1.8%
Claude CodeFable 5xhigh83.8% ± 2.3%
CodexGPT-5.5xhigh83.2% ± 2.2%
Terminus 2Fable 5high80.5% ± 2.3%

Source: Terminal-Bench 2.1, checked . The ± values show 95% confidence intervals. These are selected official entries, separate from the IndusAGI-reported campaigns above. IndusAGI was not listed in the official table checked; no official rank is claimed. Scoring and run configurations may differ.

The complete stack

Five capability layers.
One import.

Drop in the whole stack or reach for a single layer. The same architecture ships in TypeScript, Python, and Rust.

Layer 02

Agent Runtime

The model conversation loop — streaming, tool calls, and turn orchestration.

Read the documentation
agent.ts
import { runtime, capabilities } from "indusagi";

const agent = runtime.createAgent({
  model: "claude-sonnet-4",
  tools: capabilities.toolBox("coding"),
  // optional: system, maxOutputTokens, thinking, compaction, maxTurns
});

const unsubscribe = agent.subscribe((event) => {
  if (event.kind === "text_delta") process.stdout.write(event.delta);
});

const snapshot = await agent.submit("explain this repo");
console.log(snapshot.phase); // "settled" or "faulted"
unsubscribe();
INDUSAGITypeScript framework
From idea to terminal

Start building
in one command.

Free models included. No account, no credit card.

TypeScriptPythonRust
Start building in one command
$ npm install -g indusagi

Requires Node 20+ (npm), Python 3.11+ (pip), or Rust 1.96+ (cargo). Set a provider key, then run.

Live on npm

Real download stats

Pulled live from the npm registry.

SDK Packageindusagi
—per year
—monthly
—weekly
Loading stats…
CLI Packageindusagi-coding-agent
—per year
—monthly
—weekly
Loading stats…
Documentation

Pick your edition.

Framework or coding agent, in the language your team already ships.

TS

TypeScript

Framework38 pages

indusagi

The framework — LLM gateway, agent loop, TUI widgets.

Open docs
Coding agent30 pages

Coding Agent CLI

Terminal-first agent — extensions, skills, RPC & SDK modes.

Open docs
PY

Python

Framework24 pages

indusagi

The Python framework — facades over the full stack.

Open docs
Coding agent27 pages

induscode

Python coding agent — Textual console, branchable sessions.

Open docs
RS

Rust

Framework20 pages

indusagi

100% Rust, zero-SDK — compiled to a single binary.

Open docs
Coding agent27 pages

induscode

Native ratatui REPL — slash commands, MCP, capability deck.

Open docs
Frequently asked

Questions,
answered.

An open-source developer stack: a multi-provider framework plus a terminal-first coding agent. It gives developers programmatic control over autonomous agent loops, memory, tools, and the console UI.

Open source · MIT licensed

Start building
in one command.

Free models included. No account, no credit card.

Get started
npm install -g indusagi