Best Self-Hosted AI Agent Projects on GitHub
TL;DRSelf-hosted AI agent frameworks eliminate remote token markups, protect local session credentials, and provide deterministic execution sandboxes.Multi-agent engines like CrewAI, LangGraph, and AutoGen orchestrate specialized agent graphs across Python runtimes.Browser-native agents like Browser-Use and Drevon automate web tasks by pruning DOM trees or driving authenticated desktop browser sessions.Containerized engines like OpenHands isolate autonomous code execution inside Docker sandboxes to protect host file systems.
Self-hosted agent repositories now provide the primary alternative to cloud-hosted agent platforms that impose per-credit pricing and centralized credential storage. We built Drevon as an evidence-backed prospect research desktop application that runs locally on macOS, and users can download Drevon for free on Mac to execute browser-native tasks on their own hardware. Running agents locally shifts control over model routing, API budgets, and session security back to the developer.
The Shift Toward Self-Hosted and Local-First AI Agents
Self-hosted AI agents execute on developer infrastructure rather than vendor-managed cloud runtimes. This architecture avoids the credit markups common in SaaS platforms, removes third-party data retention risks, and allows direct integration with local operating system processes, local databases, and authenticated browser sessions.
Cloud-hosted agent services wrap API calls in multi-tenant queue architectures. While convenient for simple workflows, this model introduces predictable operational friction. First, third-party platforms bill for discovery through credit consumption, a model we examined when analyzing how credit-based pricing models penalize discovery. Second, routing internal company credentials or member-only web sessions through external cloud instances creates unnecessary exposure surfaces. Third, developers cannot easily customize execution loops when runtimes are managed remotely.
Evaluating open-source repositories requires assessing four technical dimensions: sandbox isolation mechanisms, DOM and context management efficiency, integration interfaces (CLI, API, or visual UI), and compatibility with local or self-hosted LLM endpoints. Managing these boundaries effectively minimizes the integration tax of AI prospecting and custom workflow engineering.
Top Self-Hosted Multi-Agent Frameworks
Multi-agent frameworks coordinate discrete LLM instances configured with specific system roles, tool sets, and execution boundaries. These engines allow developers to structure complex tasks into hierarchical or graph-based pipelines that run on local servers or personal workstations.
Three frameworks represent the current standard for code-first multi-agent orchestration:
1. CrewAI
CrewAI focuses on role-playing agent architectures where developers assign concrete goals, backstories, and task assignments to individual workers. Orchestrations execute through sequential or hierarchical processes managed by a coordinator agent. With approximately 58,100 GitHub stars and an active weekly release cadence, CrewAI integrates cleanly with custom Python tools and structured output parsers. It excels at procedural research and content transformation pipelines where task boundaries are clearly defined.
2. Microsoft AutoGen
Microsoft AutoGen pioneered conversational multi-agent architectures where agents communicate via structured message buses to resolve problems collaboratively. The repository has accumulated roughly 60,800 GitHub stars. In 2026, Microsoft began consolidating AutoGen into the broader Microsoft Agent Framework, placing the original repository into stable maintenance mode for v0.4 releases while continuing security patches. AutoGen remains an established choice for event-driven agent conversations and human-in-the-loop debugging workflows.
3. LangGraph
Maintained as part of the LangChain ecosystem, LangGraph structures multi-agent coordination as stateful cyclic computational graphs. Unlike linear execution chains, LangGraph allows agents to loop, persist intermediate state checkpoints in local databases, and branch conditionally based on tool call outputs. This makes it suitable for engineering deterministic multi-agent applications where state recovery and strict execution paths are required.
Top Browser-Native and Autonomous Task Agents
Browser-native and autonomous task agents interact directly with graphical user interfaces, terminal environments, and live web pages. Rather than relying solely on static API integrations, these tools navigate interactive environments to complete operational tasks.
1. Browser-Use
Browser-Use is an open-source web automation library with over 112,000 GitHub stars. Available via the browser-use repository on GitHub, the project connects language models to web browsers using the Chrome DevTools Protocol and Playwright.
Standard headless Playwright scripts dump raw HTML or accessibility trees, which routinely generate 10,000 to 100,000 tokens on dense web applications. Browser-Use solves this through a dedicated DOM-pruning pipeline: it strips non-interactive containers, layout noise, and invisible elements, assigning compact sequential integer tags to clickable elements (the Set-of-Marks approach). This condenses page state into 1,000 to 3,000 tokens per interaction. To prevent memory exhaustion across multi-step flows, Browser-Use retains full DOM state only for the active step, replacing older turns with action summaries.
2. OpenHands
OpenHands (formerly OpenDevin) is an autonomous software engineering agent with roughly 84,600 GitHub stars. It accepts engineering specifications, analyzes codebases, edits files, and executes terminal commands to resolve issues. According to project maintainers, approximately 20% of the project repository commits are authored by OpenHands instances resolving open pull requests.
To prevent unauthorized modifications to host files, OpenHands isolates runtime execution using container sandboxes. The architecture separates the planning loop on the host from an ActionExecutionServer inside Docker. The container mounts only explicitly declared workspace paths (`SANDBOX_VOLUMES`), runs as an unprivileged user matching host permissions (`SANDBOX_USER_ID`), and does not mount the host Docker socket, preventing container-escape risks.
3. Drevon
Drevon takes a desktop-native approach specifically for prospect research and go-to-market workflows. Instead of spinning up cloud containers or headless browsers that trigger anti-bot protections, Drevon runs locally on macOS as an Electron desktop application. We detailed this architecture when explaining why Drevon runs on your desktop instead of the cloud.
Drevon drives the user's existing authenticated browser sessions across sources like LinkedIn, Reddit, and Crunchbase. Every finding outputs to local CSV and Markdown files with inline source links verifying the claim. By connecting directly to the user's existing AI subscription (Claude, OpenAI, Gemini), it eliminates third-party data contracts and credit billing entirely.
Self-Hosted Agent Platforms with Visual Workflow Interfaces
Visual agent platforms provide web-based graphical user interfaces for designing, testing, and deploying multi-agent workflows. These platforms combine visual node editors with integrated retrieval-augmented generation (RAG) engines and multi-tenant user management.
1. Dify
Dify is the most-starred self-hosted LLM application platform on GitHub, with over 154,400 stars recorded on Star History. Maintained by LangGenius on GitHub, Dify provides visual multi-agent workflow orchestration, integrated vector database ingestion, prompt debugging interfaces, and an active Dify plugins ecosystem.
Deployable via Docker Compose, Dify allows teams to construct complex agent logic without writing boilerplate orchestration code. Its community resolves integration updates actively via Dify repository issues, making it an enterprise-grade platform for internal knowledge bases and conversational support agents.
2. Flowise
Flowise provides an open-source, drag-and-drop user interface for building LLM pipelines and agent graphs using LangChain components. It packages custom tool definitions, memory stores, and vector embeddings into composable visual blocks. Teams use Flowise for rapid prototyping before writing custom production code, running the entire stack locally via Docker or Node.js.
Comparative Architectural Breakdown
Selecting an agent framework requires aligning runtime architecture with your deployment environment, technical expertise, and security requirements. The table below compares the primary self-hosted repositories.
| Repository | GitHub Stars | Execution Runtime | Browser Control | Primary Use Case |
|---|---|---|---|---|
| Drevon | Desktop App | Local macOS Binary | Native Browser Sessions | Evidence-backed prospect research |
| Dify | ~154k | Docker Containers | API / Webhook Only | Visual RAG & enterprise workflows |
| Browser-Use | ~112k | Python / Playwright | DOM Pruning + CDP | Automated web navigation |
| OpenHands | ~84.6k | Docker Sandbox | Containerized Headless | Autonomous software development |
| AutoGen | ~60.8k | Python Process | Custom Tooling | Conversational multi-agent systems |
| CrewAI | ~58.1k | Python Process | Via Crew Tools | Role-based task automation |
Engineering teams that need custom workflow automation often adopt different tools across stages. A technical GTM engineer using a code-first approach might write custom extraction scripts in LangGraph or Browser-Use, while using Drevon for live account research without writing scraping scripts. Reviewing workflows run by AI agents shows that matching the execution boundary to the task prevents runtime failures.
Local LLM Serving vs. Hosted API Endpoints
Running multi-agent orchestrators against local models (via Ollama or vLLM) versus hosted endpoints (such as Claude 3.5 Sonnet) involves clear memory and latency trade-offs.
Multi-agent frameworks pass expanding conversation histories across iterative loops, causing Key-Value (KV) cache memory to grow rapidly. For a model like Llama 3.1 70B, KV cache overhead scales from 1.3 GB at a 4,000-token context to 42.9 GB at 128,000 tokens for a single stream. While hosted APIs offload KV cache management to cloud providers, local runtimes require sufficient VRAM to prevent Out-of-Memory crashes.
Latency also compounds in iterative agent loops. In a controlled study on concurrent LLM serving documented on arXiv (2511.21398), Ollama experienced Time to First Token (TTFT) degradation between 54 and 122 seconds under heavy concurrent request queues. Hosted endpoints like Claude 3.5 Sonnet provide steady throughput (60 to 80+ tokens per second) and support fast parallel agent branches via prompt caching. Teams balancing data sovereignty against execution speed often run lightweight planning agents locally while routing heavy analytical steps to hosted endpoints.
Frequently Asked Questions
What hardware is required to run self-hosted AI agents effectively?
Running agent orchestration frameworks like CrewAI, LangGraph, or Browser-Use requires minimal local compute (4 CPU cores, 8 GB RAM) when connecting to cloud LLM APIs. If you serve models locally using Ollama, quantised 8B models require at least 8 GB of VRAM, while 70B models require 48 GB to 80 GB of VRAM to handle multi-turn KV cache expansion without crashing.
How do self-hosted browser agents handle authentication and CAPTCHAs?
Headless frameworks like Browser-Use automate logins via script parameters or persistent browser context directories, though bot detection systems frequently challenge headless instances. Desktop-native agents like Drevon bypass this issue by operating inside the user's existing, authenticated browser session on macOS, inheriting active session cookies and avoiding bot flags entirely.
What are the primary security risks of running autonomous code-execution agents locally?
Autonomous agents that execute bash commands can accidentally delete system files, expose environment variables, or install unvetted dependencies. Projects like OpenHands address this risk by confining command execution to isolated Docker containers with explicit volume mount restrictions, non-root user permissions, and detached Docker daemon sockets.
To run evidence-backed research agents on your local workstation without data contracts or credit meters, download Drevon for Mac and start discovering qualified accounts directly through your own browser.