AI repos for aI engineers
Open-source AI projects worth it for people who build products on LLMs: agents, RAG, prompts, evals, inference serving, each with what it does for you. 180 repos, most relevant first.
-
firecrawl
Get clean markdown or schema-based JSON from live websites to feed RAG pipelines and agents.
-
markitdown
Turn folders of documents into Markdown before indexing them for RAG or prompting.
-
skills
Reference examples and spec for writing SKILL.md instructions that Claude loads per task.
-
transformers
Run local models for chat, speech recognition and vision behind your own applications.
-
browser-use
Let an LLM agent drive a real browser to complete multi-step web tasks from Python.
-
crawl4ai
Convert web pages into clean Markdown for RAG indexes or as an agent's web-reading tool.
-
headroom
Compress tool output, logs and JSON via library or proxy to cut LLM token costs without rewriting code.
-
system_prompts_leaks
Read and compare real vendor system prompts for tools, safety and formatting to inform your own agent prompts.
-
litellm
Swap LLM providers behind one OpenAI-style API and track spend, keys and limits.
-
hindsight
Add long-term memory to agents and chatbots through a self-hosted server with client libraries and integrations.
-
OpenViking
Store agent memory, docs and skills as an inspectable, editable virtual filesystem instead of an opaque vector store.
-
supermemory
Per-user memory and profile API for agents, with integrations for Vercel AI SDK, LangChain and Mastra.
-
composio
SDK giving agents authenticated tools for 1000+ apps with per-user auth, avoiding hand-written integrations.
-
openai-agents-python
Python SDK for agents with handoffs, guardrails, tracing, sessions and sandboxed workspaces.
-
dexter
Study a planning, tool-using, self-checking agent in TypeScript with inspectable scratchpad logs.
-
ai
Call many model providers through one TypeScript API for streaming, structured output and tool-calling agents.
-
12-factor-agents
Check an agent design against twelve principles for context, errors and state before shipping.
-
claude-context
Add embedding-based code search over MCP so coding agents retrieve only relevant code and use fewer tokens.
-
PixelRAG
Build RAG over screenshot tiles of pages and PDFs so tables, charts and layout are not lost to text parsing.
-
honcho
Add long-term per-user memory to LLM agents and chatbots through Python or TypeScript SDKs.
-
eve
TypeScript framework for durable agents, with tools, channels and schedules laid out as conventional folders.
-
AutomationBench
Compare LLMs on business workflows across simulated CRM, inbox and calendar tools.
-
leanctx
Compress long prompts, old chat turns and tool output before sending to OpenAI, Anthropic or Gemini, cutting token costs while code stays exact.
-
superpowers
Ready-made skills enforcing brainstorm, plan, test-first and review workflow in coding agents, useful as a pattern for agent workflows.
-
ECC
Large set of agents, skills, hooks and memory to study or reuse for coding agent workflows.
-
hermes-agent
Study an agent with persistent memory, learned skills and provider switching, plus scheduled tasks via chat.
-
n8n
Build agents on a visual canvas with your own data, tools and human approval steps.
-
open-webui
Ready chat front end with RAG, web search and plugins for Ollama and OpenAI-compatible models.
-
graphify
Builds a local tree-sitter knowledge graph instead of a vector store, so agents query structure.
-
pi
Minimal TypeScript agent toolkit with one API across LLM providers, extensions and SDK for building agent apps.
-
caveman
Cut token spend with a terse-answer rule, a proxy shrinking logs and analysis of costly agent habits.
-
TradingAgents
Reference multi-agent design with analyst, debater and risk-manager roles across many LLM providers.
-
whisper
Drop-in speech-to-text for voice features, with model sizes to trade speed against accuracy.
-
claude-mem
Hooks and MCP-based memory capture, compression and injection, a concrete pattern for agent session memory.
-
Agent-Reach
Installs and health-checks the CLI tools agents need to read YouTube, Reddit, Twitter/X and GitHub.
-
autoresearch
Shows autonomous agent experimentation steered by a single program.md file.
-
ruflo
Swarm coordination, persistent vector memory and hooks wrapped around Claude Code and Codex.
-
codegraph
Gives coding agents a local code graph over MCP, cutting tool calls and tokens when working in large codebases.
-
OpenManus
Study a small open agent framework wiring an LLM to browser tools and MCP, and experiment with it.
-
CLI-Anything
Generate CLIs so coding agents can drive desktop and web software, and install community ones from CLI-Hub.
-
multica
Assign issues to several agent CLIs on one board and compare their runs and costs.
-
codebase-memory-mcp
Study and use a code knowledge graph exposed as MCP tools to cut agent token use.
-
agent-browser
Give agents a shell-driven browser using accessibility snapshots and element refs.
-
openhuman
Configurable local-first agent harness with pluggable LLM, memory and search, embeddable in Rust.
-
financial-services
Reference agents, skills and connectors for finance workflows on Claude plugins and Managed Agents.
-
freellmapi
Single OpenAI-style endpoint that pools free-tier keys with failover for prototyping and model comparison.
-
UI-TARS-desktop
Agent stack for GUI control with vision models and MCP tools, usable with your own provider.
-
playwright-mcp
Let LLM clients drive a real browser through Playwright accessibility snapshots over MCP.
-
skills
Install and manage SKILL.md files across Claude Code, Codex and Cursor with one CLI.
-
WeKnora
Self-hosted RAG and agent platform with swappable models and storage, ready to deploy for document Q&A.
-
OfficeCLI
Lets agents create, edit and render Word, Excel and PowerPoint files and check their own output.
-
context-mode
MCP server that sandboxes tool output and persists session state, keeping agent context small across compactions.
-
gbrain
Self-hosted agent memory that stores sourced facts and answers with citations over MCP.
-
TencentDB-Agent-Memory
Shared memory server for chat memory, skills, docs wiki and code graph across Claude Code, Codex and other agents.
-
agentmemory
Local memory server shared across Claude Code, Cursor and Codex through hooks, MCP or REST.
-
symphony
Read SPEC.md and the Elixir reference to build board-driven autonomous coding agent orchestration.
-
ralph
A simple loop pattern for running coding agents unattended against a prd.json of small stories with tests.
-
jcode
Rust terminal agent harness with a shared daemon, cross-session memory and swarm workers to study or run.
-
semantica
Adds knowledge graphs and PROV-O decision audit trails so agents can explain their decisions.
-
plugins
Official reference for the Cursor plugin format, skills and MCP integrations to copy for your own.
-
openui
Lets agents answer with interactive UI using fewer tokens than JSON.
-
liteparse
Parse documents for RAG and agents, with bounding boxes for citing source locations.
-
awesome-claude-skills
Find ready-made Claude Skills and see how others structure SKILL.md files before writing their own.
-
improve
Use a capable model to audit a repo and write plans a cheaper model can execute, cutting costs.
-
sandcastle
Script coding-agent workflows, like implement-then-review pipelines, inside sandboxes with one run() call.
-
openrig
Define a team of Claude Code and Codex agents in YAML with roles and shared context.
-
plugins
Reference Codex plugins showing how manifests, skills, MCP config and agents fit together for building your own.
-
GLM-5
Open coding and agent model you can self-host with vLLM or SGLang, or call through the Z.ai API.
-
treg
One base URL and token gives agents paid third-party APIs, with team credentials kept hidden from the agent.
-
maka
Local agent runtime that logs every model call, tool call and permission decision, so runs can be replayed and inspected.
-
rakazo
Self-hosted persistent bots with memory, routines and a sandboxed browser and terminal, on a model of your choice.
-
cli-printing-press
Generates Go CLIs and MCP servers for APIs, backed by SQLite, so agents use them with fewer tokens.
-
agent-client-protocol
JSON-RPC protocol and SDKs for making a coding agent work in any ACP-supporting editor.
-
basic-memory
Persistent Markdown memory exposed through MCP, so Claude, Codex or Cursor recall project context across sessions.
-
Switchyard
Routes each agent request to a cheaper or stronger model, with benchmarks against a single-model baseline.
-
billion-context
Use a proxy that summarizes and restores old chat history so long agent sessions avoid context limits.
-
HyperQwen
Run a private OpenAI-compatible Qwen endpoint on a home GPU for batch and single-user workloads.
-
StartLux-Decision
Get calibrated probabilities per option for routing, classification and agent decision steps over an HTTP server.
-
OpenTag
Fork a working Slack and Teams agent built on AG-UI and LangGraph, then swap in your own agent.
-
channels-sdk
Put an existing AG-UI agent into Slack and Teams with buttons and approval prompts.
-
agentcookie
Keep remote browser agents logged in by syncing Chrome cookies from your Mac over Tailscale.
-
token-optimizer-mcp
Cut repeated file-read tokens in coding agents and track savings in a local dashboard.
-
peerd
Run a browser-based agent loop with your own key, tab control and sandboxed code execution.
-
cli
Give agents access to Gmail, Slack, Stripe and CRMs without building OAuth flows.
-
OSWorld-V2
Evaluate computer-use agents on long multi-step desktop and web tasks reproducibly.
-
openai-realtime-meeting-assistant
Worked example of connecting WebRTC audio to the OpenAI Realtime API with function calling to edit shared state.
-
sdk
Docs and agent skills to let agents run actions in thousands of apps without handling OAuth tokens.
-
connectors
Per-app connectors usable as agent skills or MCP tools, giving agents app access with managed auth.
-
magic-compact
Compress long Claude Code or OpenCode sessions by summarizing old assistant turns and pruning tool output while keeping user messages.
-
rta-smriti-brain
Index a repo and agent sessions into local SQLite and build evidence-labelled context packs for agent memory.
-
cortex_suite
Local MCP servers giving coding agents live code structure and learned project memory, with no API keys.
-
Iron-Jarvis
Study a local-first multi-agent setup with sandboxed git worktrees, approval gates, and Ollama or vLLM backends.
-
openclaw
Study a self-hosted agent gateway with swappable model providers, memory and multi-channel chat integration.
-
rea
Example MCP server exposing reverse engineering tools to coding agents.
-
skills
Small editable skills for TDD, diagnosis and review to copy and adapt for your own agent workflow.
-
opencode
Open source agent to use or inspect, with swappable model providers.
-
agent-skills
Examples of well-structured Markdown skills for steering coding agents consistently.
-
LLMs-from-scratch
Understand how transformers work under the hood to reason better about LLM behavior.
-
paperclip
Orchestrate many coding agents with budgets, approvals, scheduled jobs and per-run inspection.
-
orca
Compare Claude Code, Codex and OpenCode on the same task and keep the best result.
-
playwright
Reliable browser control layer for building browser-using AI agents.
-
odysseus
Compare models blind on your own prompts and try agents and deep research in one app.
-
Scrapling
Give agents a web-fetching tool through the MCP server, including stealth fetching.
-
rtk
Shrinks git, test and linter output so coding agents waste less context.
-
VoiceStudio
Local text-to-speech, voice cloning and transcription stack to study or build on.
-
ai-engineering-from-scratch
Build agent loops and MCP servers from scratch to understand what frameworks hide.
-
free-claude-code
Route Claude Code or Codex to free or cheap models across many providers and compare results.
-
voicebox
Local TTS engines with an API and MCP let you add speech to agents.
-
awesome-claude-code
Find Claude Code skills, plugins, hooks and orchestration tools to extend your agent setup.
-
CLIProxyAPI
Expose Codex or Gemini CLI logins as OpenAI, Gemini or Claude-compatible endpoints for other tools.
-
scientific-agent-skills
177 documented science skills show how to package domain tools and databases for agents.
-
DeepTutor
Reference full-stack RAG tutoring app with document knowledge bases and progress memory.
-
open-notebook
Open-source RAG notebook supporting many providers including local models.
-
airllm
Try very large open models on modest hardware via a familiar generate call.
-
VoxCPM
Apache-licensed multilingual TTS to build voice features into products.
-
Vane
Reference self-hosted answer engine combining SearxNG search, LLM citations and local models.
-
buzz
Workspace where agents are identifiable teammates in channels on a Nostr relay.
-
DeepSeek-Reasonix
Go coding agent with plan mode, sandbox, checkpoints and ACP to study or run on DeepSeek.
-
book-to-skill
Shows how to split a book into chapter-indexed skills so an agent loads only the needed content.
-
codex-plugin-cc
Adds Codex reviews and task delegation inside Claude Code for second-model opinions on code.
-
openclaude
Terminal coding agent with MCP that switches between OpenAI, Gemini, Ollama and other backends.
-
claude-code-templates
Catalog of Claude Code agents, hooks and MCP setups to reuse or learn from.
-
CloakBrowser
Gives browser-use agents a browser less likely to be flagged by bot checks.
-
cli
Gives agents Workspace access through a CLI with shipped agent skills, no custom integration code.
-
knowledge-work-plugins
Reference examples of file-based role plugins with skills, commands and tool connections.
-
9router
Local gateway routing coding tools across 40+ providers with fallback, account rotation and output compression to save tokens.
-
qmd
Local hybrid search with re-ranking over markdown and transcripts, exposed to agents through MCP.
-
modular
Serve models behind an OpenAI-compatible endpoint with the MAX inference framework.
-
page-agent
See how an in-page LLM agent reads and operates the DOM from plain-language instructions.
-
t3code
Control Claude Code, Codex and other agents from one interface, including from a phone.
-
claude-skills
Reuse ready-made skill packs and study how skills are structured for several coding agents.
-
cmux
Run several terminal coding agents in panes and see which one is waiting on you.
-
skills
Learn how Codex skill folders are structured and packaged by browsing example skills.
-
Unlimited-OCR
Serve an OCR endpoint to feed document text into RAG or extraction pipelines.
-
compound-engineering-plugin
Use a brainstorm, plan, work, review loop that saves lessons in the repo for later agent runs.
-
openwork
Point an OpenCode-based agent at local files with any model and share skills and MCP servers.
-
AI-Trader
Connect agents to a paper-trading platform to publish signals and see how agents behave in markets.
-
CodexBar
Shows remaining Claude Code and Codex limits and reset times so long agent runs don't stall mid-task.
-
text-to-cad
Example of a skill plugin that gives coding agents a real toolchain, producing STEP/STL files and manufacturability checks.
-
screenpipe
Gives agents searchable local screen and audio history as context, e.g. through Claude.
-
claude-video
Skill that feeds timestamped frames and transcripts to an agent, a working pattern for video understanding.
-
ego-lite
Lets coding agents run browser tasks in isolated Spaces using real logins, in parallel.
-
cli
Ships 26 agent skills so assistants can work in Lark without custom integration code.
-
needle
Local tool-calling model that avoids a large LLM for fixed action sets.
-
omi
Build custom apps reacting to live transcripts and conversation memory.
-
CubeSandbox
E2B-compatible private sandbox so existing agent code runs without a cloud service.
-
camoufox
Give browser-using agents a less detectable browser for web tasks on sites where automation is permitted.
-
cosmos
Run Cosmos checkpoints and cookbooks to generate robot task videos and driving scenes from prompts.
-
Octop
Study and run a multi-user, multi-agent assistant with web, CLI and chat-app channels like Telegram and Discord.
-
DesktopCommanderMCP
Give MCP clients terminal, file search and editing tools on a local machine, and study a practical MCP server.
-
universal-modder
Study agent skills, a CLI and a shared knowledge base that guide an agent through a research-to-test routine.
-
evolver
Turn agent error logs into structured improvement prompts and keep an auditable history of behavior changes.
-
osaurus
Run agents with memory, tools and sandboxed execution on local MLX or cloud models on a Mac.
-
clicky
Fork a working screen-aware voice assistant that answers aloud and points at screen elements.
-
ODS
Get a private local stack for inference, RAG, agents and voice to build and test against.
-
pxpipe
Proxy that renders Claude Code context as images, a concrete technique to cut input token spend on long sessions.
-
reasoning-from-scratch
Compare chain-of-thought, self-consistency and self-refinement on MATH-500 to understand how reasoning models work.
-
skills
Ready-made Claude Code skills for CLAUDE.md cleanup and scheduled agent loops, plus examples of how skills are written.
-
claude-plugins-community
Browse and install vetted community plugins for Claude Code, and see how to publish your own.
-
GLiNER
Cheap CPU-friendly entity extraction in place of calling a large LLM in a pipeline.
-
loopy
Catalog of agent loops with prompts, checks and stopping rules, plus a skill to build and audit your own.
-
claude-code-telegram
Python bot wrapping the Claude Code CLI with sessions, directory sandboxing, cost limits and webhook-driven reviews.
-
llms-txt
Understand the format agents consume to fetch clean site context for RAG and coding assistants.
-
birdclaw
Example of a read-only MCP server exposing personal data to agents.
-
proof-sdk
Let agents comment and suggest edits in shared documents with tracked authorship.
-
companies
Import pre-built multi-agent teams with org charts and skills to study or reuse on Paperclip.
-
postiz-agent
Skill and CLI bundle for giving coding agents social posting through the Postiz API.
-
obsidian-claude-code-mcp
See an in-app MCP server that lets Claude Code and Desktop read and edit a vault.
-
lumina
Reference for a local-first agent with layered memory, tools and voice running on llama.cpp or Ollama on modest GPUs.
-
hindsight
Skill that distills lasting lessons from coding sessions into persistent memory, a pattern to copy for agent memory.
-
rilable
Self-hosted example of a Claude agent that writes, builds and hosts web and iOS apps from prompts.
-
radiant
Mac app for running coding agents on cloud or local Ollama and LM Studio models with tool-call visibility.
-
DAA
See how an LLM agent investigates errors and opens PRs across multiple model providers.
-
glance-speedlab
Find where latency goes in local live-camera vision-language inference and tune it.
-
aura-code
Compare 31 model providers, including local Ollama, on the same coding task in one agent harness.
-
Logic-Loop
Manage several coding agent sessions and review digests of long unattended runs.
-
skills
Reuse 23 SKILL.md playbooks for planning, handoff, and multi-model decision review as agent workflow examples.
-
claude-dashboard
Inspect Claude Code transcripts, sessions, and spend from the local files it writes.
-
Clodex
Run GPT Codex models inside the Claude Code harness through a proxy and compare effort levels.
-
awesome-one
Use example workflows and an email triage tool as starting points for integration-driven agents.