LeAgent
A self-hostable desktop agent that plans, self-corrects, and ships 100+ offline tools with visual workflows and generative UI.
- Source repo
- vixues/LeAgent
- Stars
- ★ 244
- Last updated
- 4mo ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 50/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformOpenAI API · Claude API
- You'll need
- Typical use
- Back-office staff hand it a folder of invoices, contracts, or reports: it OCRs and classifies them, extracts structured data into Excel, and produces a Word/PPTX summary in one turn.
- Main limitation
- The default single-process, single-worker setup only suits SQLite's single writer; scaling out requires PostgreSQL plus sticky sessions because the execution registry and event bus are per-worker.
- Website
- vixues.com.cn ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-07-17latest release v1.2.7 (2026-07-17)
- Source review
- 50/100 · Major gaps
What does this agent do, and when should you use it?
LeAgent is an open-source desktop AI agent built as a modular monolith: a Python 3.11+/FastAPI backend, a React 19 + TypeScript SPA, and an Electron shell that bundles its own Python runtime. Every ingress — HTTP/SSE, WebSocket, cron, background task, or GenUI — mints an ExecutionRun and funnels through a thin facade into one think-act kernel, leagent.sdk.kernel.run_loop, which drives QueryEngine and its tool executor and checkpoints paused turns for exact resumption. It ships 100+ first-party tools across 13 categories, each auto-exposed as a typed ReactFlow node, plus Agent Skills v1.0, an MCP client, a YAML rule engine, and episodic/semantic/procedural memory. Outputs include Word/Excel/PPTX/PDF files, charts, and declarative generative-UI trees (KPI boards, slide decks, galleries) streamed inline over SSE and exportable. The default install needs zero external services — SQLite in a single process or one Docker container — with PostgreSQL and optional Milvus for scale, and DeepSeek, DashScope, OpenAI, Anthropic, Azure OpenAI, Ollama, or vLLM for models.
A request arrives from chat, an SDK call, cron, a webhook, or a workflow node, is wrapped in a single ExecutionRun (run_id, plus parent_run_id for child scopes), and executes inside run_loop → QueryEngine → tool executor, where the agent plans, calls tools, and self-corrects. Tools span document read/write with OCR and classification, web search via DuckDuckGo/SearXNG/Bing plus scraping, CSV/Excel cleaning and SQL/vector search, two-tier code execution (in-process RestrictedPython and a subprocess sandbox with rlimits and timeouts), Word/Excel/PPTX/PDF generators, charts and image processing, and image/video/3D/audio generation nodes. Tool results come back as FileRefs through a unified file layer with HMAC-signed preview and download URLs. A single WorkflowExecutor backs saved DAGs, in-chat step cards, cron runs, and agent-authored graphs, staging ready batches, running independent branches concurrently, and applying centralized retry/backoff and per-node timeouts. Generative UI trees are streamed and patched over SSE, rendered inline, and exportable to PDF or PPTX. State is split cleanly across TieredSessionStore, CheckpointStore (agent_checkpoints), and WorkflowStateStore, with structlog JSON logs, OpenTelemetry spans, and Prometheus histograms for observability.
- Back-office staff hand it a folder of invoices, contracts, or reports: it OCRs and classifies them, extracts structured data into Excel, and produces a Word/PPTX summary in one turn.
- Analysts who need cited briefings search across DuckDuckGo/SearXNG/Bing, scrape and download sources, then assemble a PDF report with charts and KPI dashboards.
- Data analysts load local CSV/Excel or query the managed database, clean and merge, run SQL and vector search, and get the findings narrated with generated charts.
- Researchers point it at a PDF in Research Paper Mode for outline extraction, faithful section and whole-paper summaries, reference and LaTeX-formula extraction, and region translation, all offline.
- Indie game or art teams turn a text brief into images, video, 3D meshes, and VFX sprite sheets through a node graph gated by a quality check and bounded self-correction loop, exported for Unity/Unreal/Godot.
- Teams that need private deployments run one Docker container, then configure cron jobs, inbound webhooks, and DingTalk/Feishu/WeChat Work outbound channels for always-on automation.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
docker compose up -d --buildnpm install && npm run devcurl -fsSL https://vixues.com.cn/install.sh | bash
- Environment variables
LEAGENT_SECRET_KEYDEEPSEEK_API_KEYVLLM_ENDPOINTLLM_OLLAMA_ENDPOINTDATABASE_URL- Config files
.env.example~/.leagent/.envdeploy/.env- Runtime
- Node.js 20+
- Install notes cover
- Docker
How do you install or deploy this agent?
Local development (the README's recommended path) needs git, uv, and Node.js 20+ or 22+:
git clone https://github.com/vixues/LeAgent.git
cd LeAgent
./start.sh # backend :7860 + frontend :5173Docker single-container deployment:
cd LeAgent/deploy
cp .env.example .env # set LEAGENT_SECRET_KEY + at least one provider key
docker compose up -d --buildManual step-by-step:
cd backend
uv sync --extra dev
uv run leagent init
uv run leagent appcd frontend
npm install && npm run devThere is also a one-line installer, and per-platform desktop installers ship with each GitHub release (they bundle the Python runtime, so no separate Python/Node/Docker install):
curl -fsSL https://vixues.com.cn/install.sh | bashSet at least one provider key, either as an environment variable or via Settings → Environment secrets in the web UI (which writes ~/.leagent/.env). LEAGENT_SECRET_KEY and one model key such as DEEPSEEK_API_KEY are the minimum; VLLM_ENDPOINT or LLM_OLLAMA_ENDPOINT lets you run fully local instead.
How do you use this agent?
After starting up, the Docker path serves API docs at http://localhost:8000/docs, while local dev puts the Vite frontend on :5173 and the backend on :7860. For a first run, add a DeepSeek key in Settings (the README calls it the most thoroughly validated provider, auto-aliased to tier1/tier2) and then give the chat a multi-step job, such as pointing it at an invoice folder and asking for OCR extraction plus an Excel and PPTX summary.
For visual flows, drag tool nodes onto the ReactFlow editor — every tool becomes an auto-generated typed Tool.<name> node — build a DAG, export it as YAML as a reusable template, and let the agent design, run, and refine it from chat.
For always-on automation, switch to PostgreSQL via env and wire up cron, webhooks, and IM channels:
# deploy/.env
LEAGENT_SECRET_KEY=<openssl rand -hex 32>
DEEPSEEK_API_KEY=<your key>
DATABASE_URL=postgresql://user:pass@host:5432/leagent
LEAGENT_DEBUG=falseA complete backup means the SQLite database (WAL mode) plus the working-uploads, knowledge, and coding-project trees under LEAGENT_HOME.
What are this agent's strengths and limitations?
- One think-act kernel serves chat, SDK, background tasks, sub-agents, and workflow nodes, with paused turns checkpointed and resumed in place and clearly owned state per subsystem.
- Every tool is auto-lifted into a typed ReactFlow node, so exposing a new capability visually needs no glue code; the WorkflowExecutor handles ready-batch concurrency, retries, and timeouts.
- Zero-config default really runs: SQLite in one process or one Docker container, with DeepSeek/DashScope/OpenAI/Anthropic/Azure or fully offline Ollama/vLLM.
- End-to-end output coverage — Office documents, PDFs, charts, generative UI that streams into chat and exports, plus citation-grounded Research Paper Mode.
- Extensible in several directions at once: Agent Skills v1.0 with progressive disclosure, an MCP client, inbound webhooks, outbound IM channels, and a declarative YAML rule engine.
- The default single-process, single-worker setup only suits SQLite's single writer; scaling out requires PostgreSQL plus sticky sessions because the execution registry and event bus are per-worker.
- You must bring model keys or a local inference endpoint, and provider depth is uneven — DeepSeek is documented as the most thoroughly validated.
- The desktop app is explicitly Beta with features still being refined, and the macOS builds are unsigned, requiring a manual
xattr -dr com.apple.quarantinestep. - First-time local dev pulls in uv, Node.js, and Playwright Chromium; the Docker path still needs keys and Compose, so setup is more than a single command.
- Vector-backed memory recall depends on optional Milvus and degrades without it, and backups must include the whole LEAGENT_HOME tree rather than a single database file.
How does this agent compare with similar options?
The README positions LeAgent against closed cloud chatbots and code-only CLIs: unlike chat-only cloud tools it emphasizes local 100+ tools, visual workflows, and generative UI; unlike code-only CLIs it folds documents, media, and a game-art asset pipeline into one runtime. It does not name any specific competing self-hosted project.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| LeAgent This agent | 50 · Major gaps | — | ★ 244 | 4mo ago | Python | OpenAI API · Claude API |
| LazyMind | 61 · Some gaps | Desktop appFree + model costs | ★ 790 | 1d ago | Go | Codex |
| NextClaw | 61 · Some gaps | Desktop appFree + model costs | ★ 264 | 1d ago | TypeScript | Codex · Claude Code |
| PocketPaw | 44 · Major gaps | Desktop appFree + model costs | ★ 934 | 2d ago | Python | Codex · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
least_privilege: security docs describe deny/allow/ask, Plan Mode, PathSandbox and per-tool check_permissions at the execution layer, but this is documentation only with no implementation code in the packet, so 2 not 3. user_confirmation: docs claim destructive-action confirmation and allow_once/allow_session/deny approvals with ambiguous replies defaulting to deny, again doc-only, capped at 2. data_flow_transparency: README claims data stays local by default but does not specify which fields leave to providers/MCP/webhooks or retention, 1. sensitive_data_handling: docs admit the logging pipeline does not prove all fields are redacted and that Trace payload capture is off by default (good default), but no verifiable implementation, 1. dependency_security: CI runs lint/test/build but ruff steps are continue-on-error and there is no dependency lock audit or vulnerability scanning evidence, 1. external_effects: sandbox, approval and rate-limit designs are described but unbacked by code, 2. rollback: checkpoints for pause/resume and backup notes exist, but no transaction rollback or undo for destructive actions, 1. source_attribution: SKILLs require citing managed file_id, but provenance for external content remains doc-level only, 1.
self_consistency: README, AGENTS.md, CLI and main.py agree on ports, execution topology and commands, 2. dependency_availability: README lists uv/npm/Playwright/Docker prerequisites but no pinned versions or offline-availability verification, 1. failure_messages: main.py has fail-fast migration and warning logs, but many paths suppress exceptions as non-fatal and user-facing error text evidence is limited, 1.
audience_and_scenarios: README targets self-hosting/privacy users and enumerates office, research and data scenarios, 2. capability_boundaries: SKILL files state explicit in/out of scope and security docs honestly bound capabilities, 2. trigger_precision: SKILL descriptions and tags give reasonably clear trigger conditions, 2. environment_fit: supports local, Docker and desktop across platforms, though macOS builds are unsigned requiring manual quarantine removal, 2.
information_architecture: README and AGENTS.md give a clear directory and subsystem map, 2. install_notes: multiple install paths (local, Docker, manual, one-line), 2. naming_stability: v1/v2 APIs, compat shims and tier migrations exist; naming/interface stability is not evidenced, 1. examples_and_faq: example commands and scenarios present but no FAQ, 1. known_limitations: README notes desktop Beta, SQLite single-writer and sticky-session needs, but scattered and unsystematic, 1. license: full Apache-2.0 LICENSE and SKILLs declare Apache-2.0, 3. versioning_changelog: release badge and SKILL version fields exist but no CHANGELOG file, 1. maintenance_responsibility: SECURITY.md gives contact and response SLAs, but publisher identity is unverified so ownership is unclear, 1.
output_usability: SKILLs require managed file_id/download_url and structured tables, giving usable outputs, 2. marginal_value: integrates tools, workflows, GenUI and offline capability, adding value over generic chat, 2. cost_benefit: broad feature surface but heavy dependencies, many config knobs and unsigned desktop builds; cost/benefit evidence is insufficient for a high score, 1.
claim_traceability: README claims point to docs paths but corresponding implementation files are not in the packet, 1. cross_source_corroboration: README, AGENTS.md and security docs corroborate the architecture, but all are self-description, 1. fact_inference_separation: security docs honestly separate implemented from unimplemented (e.g. admitting no complete injection detector), but SECURITY.md makes unbacked claims (input sanitized, CSRF protection) without code evidence, 1.
- Security controls (permissions, approvals, sandbox, RBAC) are largely documentation claims; implementation code is absent from the packet and should be audited before deployment.
- SECURITY.md asserts input sanitization, CSRF protection and compliance support without code evidence; do not treat these as verified.
- The one-line installer is served from a third-party domain (vixues.com.cn) and the publisher is unverified; review the script before executing.
- macOS builds are unsigned and require manual quarantine removal, posing supply-chain and integrity risk.
- CI ruff steps are continue-on-error and there is little evidence of dependency vulnerability scanning or lock auditing.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
[1.2.7] - 2026-07-17 Fixed - frontend: Restore pruned wasm optional deps in package-lock([23229e4](https://github.com/vixues/LeAgent/commit/23229e4de1d629cf32beb2ecf4af46219bc091d8)) - ci: Pin setup-uv to v8.3.2([2d491c8](https://github.com/vixues/LeAgent/commit/2d491c88779d4c94b2efc109e37f9a2142469351)) - tests: Restore os.environ after configure_settings apply tests([a046059](https://github.com/vixues/LeAgent/commit/a0460597b4e755dfb0bafd78983261ca3d77d597)) …View on GitHub ↗ -
Release notes excerpt
[1.2.6] - 2026-07-16 Added - Introduce Agent SDK/runtime with checkpoint resume and unified loop([8ce403f](https://github.com/vixues/LeAgent/commit/8ce403f0256170a884a85370db31f64453e46ce8)) - Add domain-model registry with workflow nodes and FlowPage refactor([5a8d792](https://github.com/vixues/LeAgent/commit/5a8d792878a1f99132514e93dbb221e7c34c8093)) - backend: Harden workflow executor resume, outputs, and event bus([47a2445](https://github.com/vixues/LeAgent/commit/47a244500eee0992bc04a27b8fc6978ec6ea92ce)) …View on GitHub ↗ -
Release notes excerpt
[1.1.4] - 2026-06-05 Added - Enhance code execution and project editing tools([b83de45](https://github.com/vixues/LeAgent/commit/b83de45cdf657414a8f20203df81b486ccc243f1)) Changed - Canvas HTML serving and Playwright asset signing([3882f90](https://github.com/vixues/LeAgent/commit/3882f90b4cebbffbeb6edcd8997f28156b34fe2b)) Full Changelog: https://github.com/vixues/LeAgent/compare/v1.1.3...v1.1.4View on GitHub ↗
FAQ
Do I need a cloud model, or can it run offline?
Where does my data live?
Can it handle a team or production load?
What happens when a turn needs my input?
How mature are the desktop installers?
Is LeAgent open source?
Is LeAgent safe to use?
Is LeAgent still maintained?
Who makes LeAgent?
What is LeAgent's official website?
Thanks — we use this to improve the page.