Move your AI agent
onto niubash

Windows is the first battleground: every command the agent runs lands on real bash syntax, with deterministic output and git 3–7× faster. Agents are grouped into terminal / VSCode family / workarounds / not supported / not yet verified; cards with a switch script open to a copy-paste script, verification and a rollback path.

Wire up your AI agent →

24 AI agents catalogued — 15 work out of the box · 38 more in the research pipeline

① I use Claude Code / Codex / Cursor…

Pick your agent in the directory — terminal-native and VSCode-family entries ship a switch script, verification and a rollback path.

Browse the directory →

② My agent isn't listed

The universal wiring guide: four detection mechanisms + a three-step decision flow; any agent can be wired to niubash.

Read the universal guide →

③ Let the agent install it itself

The AI portal: paste the one-click install prompt and your agent installs and wires up niubash on its own.

Open the AI portal →

Works out of the box (terminal)

6 Wire the agent to niubash today using the methods below

VSCode family

9 One shared probe, one entry each for all nine: VS Code and its forks / extensions — fix it once, the whole family follows
VS Code(Copilot agent) Ready
IDE

The family anchor: the GitHub Copilot agent runs commands through the editor's automation terminal and hits the family-shared probe by default — no configuration needed.

VS Code family probe / automationProfile one-line script ✓ open →
Windsurf Ready
IDE

A Code OSS lineage fork: the automation terminal inherits the family-shared probe — it hits the launcher once niubash is installed, no configuration needed.

VS Code family probe / automationProfile one-line script ✓ open →
Trae Ready
IDE

A VS Code lineage fork: the automation terminal inherits the family-shared probe — it hits the launcher once niubash is installed, no configuration needed.

VS Code family probe / automationProfile one-line script ✓ open →
Kiro Ready
IDE

AWS's agentic IDE (Code OSS lineage): wired per the family-shared probe on an inferred basis — agents.md marks it inferred, not machine-tested; it should hit the launcher once niubash is installed.

VS Code family probe / automationProfile one-line script ✓ open →
ZCode Ready
IDE

The family's only machine-verified member: the probe sequence in agents.md was read out of a ZCode app.asar (verified locally) — hits by default, no configuration.

VS Code family probe / automationProfile one-line script ✓ open →
Cline Ready
Editor extension

An editor extension, not a standalone IDE: it never picks a shell itself — commands run through the host editor's terminal or automation profile and inherit the family probe.

VS Code family probe / automationProfile one-line script ✓ open →
Roo Code Ready
Editor extension

A Cline-lineage editor extension: it never picks a shell itself — commands run through the host editor's terminal or automation profile and inherit the family probe.

VS Code family probe / automationProfile one-line script ✓ open →
Kilo Code Ready
Editor extension

An editor extension (Cline/Roo lineage): it never picks a shell itself — commands run through the host editor's terminal or automation profile and inherit the family probe.

VS Code family probe / automationProfile one-line script ✓ open →
Continue Ready
Editor extension

An editor extension: it never picks a shell itself — commands run through the host editor's terminal or automation profile and inherit the family probe.

VS Code family probe / automationProfile one-line script ✓ open →

Workarounds

3 Doable in some scenarios; waiting on upstream

Not supported yet

2 Upstream exposes no mechanism; issues are being tracked

Not yet verified

4 No per-machine test or code read yet — wire via the universal guide and report back

Research pipeline · 调研中

38 Candidates from the ecosystem survey, none verified yet — one gets catalogued each time one is verified

24 agents catalogued (15 work out of the box); another 38 candidates are in the research pipeline — priority goes to CLI/TUI agents that execute a local shell.

Top 10 · Priority

Other CLIs & products 9 Candidate products beyond the terminal-first set

NameFormVendorOutlookOne-liner
Open Interpreter ↗ 59k★ CLI Open Interpreter Needs config 59k★ veteran project that lets LLMs run code/commands locally — the host-shell call point awaits a code read.
Auggie ↗ CLI Augment Code Needs config Augment Code's terminal agent (Auggie CLI, closed source) — awaits verification.
Codebuff ↗ CLI Codebuff Needs config A terminal coding agent with parallel multi-agent runs — awaits verification.
Grok Build ↗ Web xAI Hard to verify xAI's coding agent product, closed source with an undisclosed execution environment — too little verifiable surface.
grok-cli ↗ 3.5k★ CLI superagent-ai Needs config Community Grok terminal agent at 3.5k★ — awaits a code read.
Vibe CLI ↗ CLI Mistral AI Needs config Mistral's agentic coding CLI — awaits a code read.
ShellGPT ↗ 8k★ CLI TheR1D Needs config 8k★ natural-language-to-shell generator running locally — the wiring point is clear; the argument style awaits a code read.
aichat ↗ 9k★ CLI sigoden Ready-candidate sigoden's 9k★ all-in-one CLI: shell execution takes $SHELL — should work from a niu session; mechanism clear, machine test pending.
Forge ↗ 3k★ CLI Antinomy Needs config 3k★ open-source agent CLI in Rust — awaits a code read.

Frameworks & engines 6 The wiring point lives in user code; no single verifiable surface

NameFormVendorOutlookOne-liner
AutoGPT ↗ 187k★ Framework Significant Gravitas Hard to verify 187k★ elder of the autonomous agent frameworks — executors are assembled by user code; no single wiring point.
gpt-engineer ↗ 55k★ Framework Anton Osika Hard to verify 55k★ codegen pipeline framework — the shell is decided by user code.
MetaGPT ↗ 50k★ Framework FoundationAgents Hard to verify 50k★ multi-role 'software company' framework — the shell is decided by the execution environment.
CrewAI ↗ 60k★ Framework CrewAI Hard to verify 60k★ multi-agent orchestration framework — tool execution is assembled by the user.
smol developer ↗ Framework smol-ai Hard to verify A minimal human-in-the-loop codegen prototype — library-shaped, no fixed wiring point.
AutoCodeRover ↗ CLI nus-apr Hard to verify Research-grade autonomous bug-fixing agent (SWE-bench lineage), usually containerized — awaits verification.

IDE & editor extensions 7 Commands mostly run through the host editor terminal

NameFormVendorOutlookOne-liner
Tabnine Agent ↗ IDE plugin Tabnine Needs config In-IDE agent mode; commands run through the host editor terminal — follows automationProfile-style settings.
Sourcegraph Cody ↗ Editor extension Sourcegraph Needs config Editor extension; commands execute through the host terminal — follows the VS Code family probe / automationProfile.
JetBrains AI Assistant ↗ IDE plugin JetBrains Needs config Plugin across the JetBrains IDE family — the built-in terminal shell can point at niu.exe; whether that reaches the agent awaits testing.
Gemini Code Assist ↗ Editor extension Google Needs config Agent mode of the VS Code/JetBrains extension — its state does not transfer to the blocked Gemini CLI (unconditional PowerShell); awaits testing.
avante.nvim ↗ 15k★ Editor extension yetone Ready-candidate 15k★ Neovim plugin: Neovim's 'shell' option (&shell) can point at niu.exe — mechanism clear, machine test pending.
codecompanion.nvim ↗ 6.9k★ Editor extension olimorris Ready-candidate 6.9k★ Neovim plugin: agent commands run through Neovim's 'shell' option — mechanism clear, machine test pending.
gptel ↗ Editor extension karthink Needs config An Emacs LLM client: tool execution spawns system shells through Emacs subprocesses — awaits verification.

Chinese products 6 Standalone IDEs and editor plugins

NameFormVendorOutlookOne-liner
通义灵码 Lingma ↗ IDE plugin Alibaba Cloud Needs config Alibaba Cloud's AI coding assistant (VS Code/JetBrains plugin) — commands run through the host editor terminal.
Qoder ↗ GUI Alibaba Needs config Alibaba's agentic IDE (standalone product) — terminal detection awaits a code read / testing.
文心快码 Comate ↗ IDE plugin Baidu Needs config Baidu's AI coding assistant (VS Code/JetBrains plugin) — commands run through the host editor terminal.
CodeGeeX ↗ IDE plugin Zhipu AI Needs config Zhipu AI's coding plugin (many editors) — commands run through the host editor terminal.
iFlyCode ↗ IDE plugin iFlytek Needs config iFlytek's AI coding assistant (VS Code/JetBrains plugin) — commands run through the host editor terminal.
Fitten Code ↗ IDE plugin FittenTech Needs config FittenTech's AI coding plugin (many editors) — commands run through the host editor terminal.

Outlook legend: Ready-candidate = mechanism known (configurable shell / takes $SHELL), awaiting a machine test · Needs config = probably configurable, awaiting a code read or test · Likely blocked = blocked by design (cloud execution / sandbox / hard-coded shell) · Hard to verify = closed source or no stable verifiable surface; sources are the outbound links on each row. Survey date: 2026-10-22.

The principle: how it picks a shell ($SHELL rule + uname signature table)

Launch the agent from inside a niu session: niu exports $SHELL, and any agent that honors the POSIX convention picks it up. $SHELL is not persisted — agents started from the Start menu will not see it; configure those per their detail pages.

Check which shell any agent is actually using: echo "SHELL=$SHELL"; uname -s; uname -a

In plain terms: have the agent run the verify command above and match its output against this table — MSYS_NT means it is already on niu (success); MINGW64 means it fell into Git Bash; Linux means it fell into WSL; a missing command with an empty $SHELL means it is still on PowerShell/cmd.

uname -s = MSYS_NT-… (uname -a ends with x86_64 Msys) niu (winuxcmd shim) ✓ MSYS_NT is a deliberate compatibility persona (uname reports MSYS so old scripts assume they are inside MSYS), not a real MSYS runtime — there is no MSYS inside niu.
uname -s = MINGW64_NT-… (MSYSTEM=MINGW64) Git Bash
uname -s = Linux WSL — the agent fell back to System32\bash.exe; fix per the detail page
uname missing, $SHELL empty PowerShell or cmd

Status legend: Ready = configurable in the current version · Partial = workaround for some scenarios · Blocked = upstream exposes no mechanism · Unverified = no test or code read yet; source-read (not machine-tested) entries name their evidence in the detail page notes.

Detail pages behind these cards are English translations of the Chinese source (agents.md); the Chinese editions stay one click away on every page.

Source of truth for this page: Coding agents on Windows, and the shell they run (niubash repo docs/src/agents.md); icon provenance and licensing live in the repo's THIRD_PARTY_NOTICES.md