Pinvou Agent Desktop Workspace

Turn research, design, and coding tasks into editable deliverables in one desktop workspace.

Stars
★ 3k
Last updated
today
License
MIT
Primary language
Rust

At a glance

How it runs
Desktop app
Works with
Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Git with submodule supportNode.js and npmCurrent Rust toolchainTauri 2 system dependenciesOpenAI-compatible model endpointShell / CLINetwork accessLocal filesystemMCP Server
Typical use
A team needs to synthesize research from PDFs, Office files, and an internal knowledge base into reusable documents.
Not a fit if
  • Users who need an out-of-the-box model without configuring an endpoint
  • Teams that require all requests to stay local while using cloud models or remote connectors
Website
pinvou.com ↗ set on GitHub
Maintenance
Commits in the last 30 dayslast commit 2026-11-20latest release v1.0.2 (2026-11-18)+3.1 ★/day over 7 days

What does this agent do, and when should you use it?

Pinvou Agent is an open-source desktop app for work, design, and coding, with a React and Vite interface integrated with the operating system through Tauri 2. The desktop app delegates agent work to the CodeWhale engine, which supports OpenAI-compatible model services, local vLLM, MCP servers, CLI connectors, Skills, Commands, Hooks, and context compaction. Users can work with attachments and a knowledge base, generate and edit visuals, or run Codex, Claude Code, and Kimi through ACP against a real project or an isolated temporary workspace. Sessions, tool calls, and artifacts persist, and created or edited files are collected in an artifact panel for preview and opening. It runs on Windows, macOS, and Linux; whether data leaves the device depends on the chosen model and tools, since cloud models and remote services receive relevant requests.

Users can attach PDFs, Office documents, images, and text by dragging or pasting them into a session. The app can combine those inputs with full-text and vector retrieval from a local knowledge base, personas, Skills, MCP tools, and workflows to research, analyze, and write, then collect resulting files in the artifact panel. Design mode generates posters and data visualizations from prompts, then lets users select elements and adjust copy, fonts, colors, dimensions, and layout. Code mode connects Codex, Claude Code, or Kimi through ACP so they can read and edit a selected project, run commands, and surface plans, tool steps, permission requests, and file changes. The app calls local vLLM or OpenAI-compatible model endpoints and can connect external tools through MCP servers and CLI connectors; sessions remain bound to their workspace and can be resumed after an app restart.

  1. A team needs to synthesize research from PDFs, Office files, and an internal knowledge base into reusable documents.
  2. A designer wants to draft a poster or data visualization from a prompt and then refine its copy and layout directly.
  3. A developer wants Codex, Claude Code, or Kimi to edit a real project and run commands from a desktop workspace.
  4. An individual or small team wants to combine specialist personas, Skills, MCP tools, and workflows for recurring office tasks.
  5. A user wants to run private workflows with local vLLM and local tools.

Key setup facts

Pulled from the install and usage notes below. The project README has the current values.

Environment variables
DEEPSEEK_BASE_URL DEEPSEEK_API_KEY DEEPSEEK_MODEL
Install notes cover
Windows · macOS · Linux

How do you install or deploy this agent?

Building from source requires Git with submodule support, Node.js and npm, a current Rust toolchain, and the Tauri 2 system dependencies for the target platform. You also need an accessible OpenAI-compatible model endpoint; local vLLM is supported. The source tree supports Linux, Windows, and macOS.

How do you use this agent?

Clone the repository and start the development app from its root:

git clone --recursive https://github.com/Pinvou/pinvou-agent.git
cd pinvou-agent/pinvou3-app
npm ci
cd ..
./pinvou3-app/run-dev.sh

Example configuration for a local vLLM endpoint:

export DEEPSEEK_BASE_URL="http://127.0.0.1:8000/v1"
export DEEPSEEK_API_KEY="local-no-auth"
export DEEPSEEK_MODEL="your-model-name"

You can also manage endpoints, model names, and API keys in the app settings. After startup, create a session and select a model, then add attachments or tools. For a coding task, use Code mode to choose an ACP coding agent and project workspace.

What are this agent's strengths and limitations?

Pros
  • One desktop app covers work, editable visual design, and ACP coding flows, with generated files collected for review.
  • Supports local vLLM and any OpenAI-compatible endpoint, with saved model configurations selectable per session.
  • The knowledge base offers both full-text and vector retrieval, supports multiple collections per chat, and retains answer provenance.
  • MCP, CLI connectors, Skills, Commands, and workflows can be combined, alongside a range of documented third-party connectors.
Limitations
  • Running from source requires Git submodules, Node.js, Rust, Tauri 2 system dependencies, and a model endpoint, so setup involves several steps.
  • Cloud models, remote MCP servers, and third-party connectors send relevant requests to those services; local use requires configuring local models and tools.
  • Parsing files, OCR, and archives can depend on optional external tools; speech engines may also need to be packaged per build configuration.
  • The project is under active development; the README points to the main branch and latest release notes for current behavior, and in-app update checks are not enabled yet.

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Pinvou Agent Desktop Workspace This agent 56 · Major gaps Desktop appFree + model costs ★ 3k today Rust Codex · Claude Code · OpenAI API · Claude API
Material Learning Studio 82 · Good Agent plugin / skillFree ★ 1.5k 1mo ago Python Codex · Claude Code
DeepChat 70 · Some gaps Desktop appFree + model costs ★ 6.4k 1d ago TypeScript OpenAI API · Claude API
Atom Platform 67 · Some gaps CLIFree + model costs ★ 912 today Python OpenAI API · Claude API

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
56/ 100 5-point scale 2.8 / 5
This review covers the source as of Nov 12, 2026. The project has shipped 2 stable release(s) since, which have not been re-reviewed. The score and verdict describe the reviewed version only. Recent releases ↓
Trust 16/29
Reliability 8/14
Adaptability 12/18
Convention 10/18
Effectiveness 7/13
Verifiability 3/8
Why each dimension lost points
Trust16 / 29 · 2.8/5

least_privilege: skills and code show credentials in the OS credential store, ima tool restricted to an allowlisted api_path and fixed endpoint, workspace roots normalized with ancestor re-checks; no complete permission inventory, deducted. user_confirmation: SKILL.md requires outline confirmation, asks before ambiguous writes, Plan/YOLO modes distinguished, but YOLO allows direct execution, deducted. data_flow_transparency: README states local models/tools stay local and cloud models/remote MCP send data out, good; no per-tool data-flow inventory, deducted. sensitive_data_handling: credentials kept out of logs, no plaintext JSON, atomic writes with 0600 and backups, backed by static test assertions; relies on assertions rather than executed verification, deducted. dependency_security: only pinned connector binaries in CI and an SBOM doc mentioned; no dependency audit/lock policy evidence, scored 1. external_effects: remote-control relay, external commands, network access exist and AGENTS.md asks to inform users, but no unified external-effect inventory or default-off policy, scored 1. rollback: knowledge base has encrypted backup/restore and data-dir locks, but no general rollback for destructive file/command actions, scored 1. source_attribution: README cites THIRD_PARTY_NOTICES, SBOM, TRADEMARKS and knowledge answers retain collection/file provenance, good but no per-item attribution, scored 2.

Reliability8 / 14 · 2.9/5

self_consistency: README, AGENTS.md, SKILL.md and code agree on architecture boundaries, single version source, deferred tool loading, scored 2. dependency_availability: depends on CodeWhale submodule, external Tauri system deps, optional external parsers, pip deps users must install; fragile install path, scored 1. failure_messages: explicit Chinese error constants (data dir busy, model install busy) and SKILL.md guidance for blocked_by_validate and mcp_boot connecting, scored 2.

Adaptability12 / 18 · 3.3/5

audience_and_scenarios: README covers work/design/code scenarios and multiple languages, scored 2. capability_boundaries: SKILL.md states when to use/not use, degradation when MCP missing, forbids HTML instead of pptx, scored 2. trigger_precision: each SKILL.md description gives explicit trigger phrases though some are broad, scored 2. environment_fit: README details Linux/Windows/macOS and LoongArch support boundaries, scored 2.

Convention10 / 18 · 2.8/5

information_architecture: README well structured, AGENTS.md has policy entry points, repo layout documented, scored 2. install_notes: README has prerequisites, source run, submodule handling and points to INSTALL.md, scored 2. naming_stability: VERSION single source, sync-version script, naming rules explicit, scored 2. examples_and_faq: command examples and screenshots but no FAQ, scored 1. known_limitations: README states in-app update not enabled, old-world LoongArch unsupported, no local file upload, scored 2. license: full MIT text present, scored 2. versioning_changelog: VERSION and Releases exist but no CHANGELOG file, scored 1. maintenance_responsibility: SECURITY.md says best-effort, no SLA, no bounty; publisher identity unverified, responsibility unclear, scored 1.

Effectiveness7 / 13 · 2.7/5

output_usability: artifact cards, editable Markdown, editable pptx, compliant docx rendering; good output usability, scored 2. marginal_value: integrates tools, knowledge, design and coding in one desktop workspace, incremental over generic chat, scored 2. cost_benefit: requires Rust/Node/Tauri toolchain, submodules, optional external parsers and model downloads; high deployment cost, scored 1.

Verifiability3 / 8 · 1.9/5

claim_traceability: README claims mostly point to docs and code, but some feature claims (OAuth/SSO, remote control) lack implementation evidence in the sample, scored 1. cross_source_corroboration: README, AGENTS.md, SKILL.md and tests partially corroborate each other but the sample is incomplete, scored 1. fact_inference_separation: SKILL.md explicitly forbids fabricating document numbers, amounts, legal clauses and requires pending-review markers, good practice; but README marketing statements mix with implementation evidence, scored 1.

Risks and how to mitigate them
  • Publisher identity is not verified by the FollowAgents curated enterprise registry; maintenance responsibility and update path are unclear, verify independently before adoption.
  • Depends on the CodeWhale submodule, Tauri system dependencies and optional external parsers; the install path is fragile, confirm each item via INSTALL.md.
  • YOLO mode allows direct execution, and the remote-control relay and third-party connectors send data out; assess data boundaries before enabling.
  • Evidence for dependency security and rollback is limited; destructive file/command actions lack general rollback, so trial in an isolated environment first.
  • This is a static source review with no build or execution performed; feature claims and test effectiveness are not independently verified.
Evidence confidence: Low Reviewed Nov 12, 2026 Reviewed revision b7a814bc8b90 New commits since this review; the score may not cover them
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/pinvou-agent.svg)](https://followagents.com/en/agents/pinvou-agent)
<a href="https://followagents.com/en/agents/pinvou-agent"><img src="https://followagents.com/badge/pinvou-agent.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v1.0.2 Pinvou 1.0.2
    Release notes excerpt
    官方下载 / Official Downloads Pinvou 已在新华三官网提供官方下载渠道: - Windows x64:[新华三官网 — pinvou3 for Windows](https://www.h3c.com/cn/d_202607/2878418_30005_0.htm) - macOS Universal:[新华三官网 — pinvou3 for macOS](https://www.h3c.com/cn/d_202607/2887430_30005_0.htm) - H3C LinSeer MegaCube(Linux ARM64):[新华三官网 — pinvou3 for MegaCube](https://www.h3c.com/cn/d_202607/2877961_30005_0.htm) > 新华三官网与 GitHub Assets 均为官方分发渠道,版本同步可能存在时间差;下载前请确认页面显示的版本。v1.0.1/v1.0.2 官网页面尚未同步更新,当前版本请优先使用本页下方 Assets。 …
    View on GitHub ↗
  2. v1.0.1 Pinvou 1.0.1
    Release notes excerpt
    官方下载 / Official Downloads Pinvou 已在新华三官网提供官方下载渠道: - Windows x64:[新华三官网 — pinvou3 for Windows](https://www.h3c.com/cn/d_202607/2878418_30005_0.htm) - macOS Universal:[新华三官网 — pinvou3 for macOS](https://www.h3c.com/cn/d_202607/2887430_30005_0.htm) - H3C LinSeer MegaCube(Linux ARM64):[新华三官网 — pinvou3 for MegaCube](https://www.h3c.com/cn/d_202607/2877961_30005_0.htm) > 新华三官网与 GitHub Assets 均为官方分发渠道,版本同步可能存在时间差;下载前请确认页面显示的版本。v1.0.1 官网页面尚未同步更新,当前版本请优先使用本页下方 Assets。 …
    View on GitHub ↗

FAQ

Can I run it entirely on my own machine?
Yes, a local model with local tools can keep the workflow fully local. If you enable cloud models, remote MCP servers, or third-party connectors, relevant requests go to those services.
Am I locked to one model provider?
No single provider is required. The app supports local vLLM and OpenAI-compatible APIs, includes templates for several providers, and allows custom compatible endpoints.
What can an ACP coding agent access and change?
With Codex, Claude Code, or Kimi, you can select a real project or an isolated temporary workspace. The agent can read and edit project files and run commands; the app surfaces tool steps, permission requests, and file changes.
Are ready-to-install packages available?
The README links to GitHub Releases and states support for Windows, macOS, and Linux. Running from source requires Git, Node.js/npm, Rust, Tauri 2 system dependencies, and a configured model endpoint.
Is Pinvou Agent Desktop Workspace free?
The software is free, but Pinvou Agent Desktop Workspace calls an AI model, and you pay that model provider for usage (or run a local model).
Is Pinvou Agent Desktop Workspace open source?
Yes. The source code is on GitHub (Pinvou/pinvou-agent) under the MIT licence.
Is Pinvou Agent Desktop Workspace safe to use?
FollowAgents' FARS review (Nov 12, 2026) scored Pinvou Agent Desktop Workspace 16/29 on trust and safety. None of the eight safety controls the review checks for was missing outright. The repository has changed since that review.
Is Pinvou Agent Desktop Workspace still maintained?
Yes. The latest commit on GitHub is from 2026-11-20, within the last 30 days. The latest release is v1.0.2 (2026-11-18).
Who makes Pinvou Agent Desktop Workspace?
Pinvou Agent Desktop Workspace is published on GitHub by the account Pinvou. FollowAgents has not verified which person or company is behind that account.
What is Pinvou Agent Desktop Workspace's official website?
The project's GitHub repository lists https://pinvou.com/ as its website. That link is set by the repository owner; the source code itself is on GitHub (Pinvou/pinvou-agent).
Was this page helpful?
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related concepts