GitHub Projects for AI Engineers
190 repositories that keep resurfacing in the AI engineering conversation, organized by the part of the stack they are changing.
- Projects
- 190
- Combined stars
- 8.1M
- Combined forks
- 975.6K
- Theme lanes
- 5
GitHub metrics checked 2026-09-13. Ranked by stars, then forks. Growth is a dated GitHub Trending observation; missing values are unobserved.
Popularity skyline
The ten most-starred projects in this edition. Height represents GitHub stars.
One stack, two ways to read it
Browse by capability when choosing components, or by production depth when diagnosing what sits underneath a working AI feature. Every item opens an official repository, product page, or documentation route.
- Tools
- 98
- Capability lanes
- 13
- GitHub links
- 65
- Official links
- 33
This is a selection map, not a recommendation to use every layer or tool.
Deployment & inference
Run models and AI workloads with an explicit serving, cloud, latency, and cost boundary.
LLMs
Provide the model intelligence layer; model choice is one dependency, not the whole system.
Agentic orchestration
Coordinate state, planning, tool calls, handoffs, and multi-agent execution.
RAG frameworks
Turn retrieval, grounding, ranking, and generation into inspectable pipelines.
Embeddings
Translate content into a semantic space for search, clustering, and retrieval.
Data extraction
Convert web pages, documents, and PDFs into structured, attributable context.
MCP
Expose tools and context sources through a shared protocol rather than client-specific glue.
AI security
Constrain unsafe input, output, data exposure, and tool execution before production writes.
Evaluation & observability
Trace behavior, measure quality, compare versions, and catch regressions across the full loop.
Memory
Persist state and knowledge beyond one context window while keeping updates auditable.
Agent SDKs
Package tools, handoffs, state, and runtime behavior into application-ready agent components.
Automation
Connect model decisions to durable workflows, retries, schedules, and human approval gates.
Vector databases
Store and query semantic representations with metadata, filtering, and operational controls.
Deployment
Where the workload runs and how inference reaches production.
Evaluation
How quality, traces, retrieval, and regressions are measured.
LLMs
The reasoning and generation engines selected for each task.
Framework
The orchestration and retrieval layer that binds models to application logic.
Vector DB
The semantic retrieval substrate for indexed knowledge.
Embeddings
The representation layer that controls semantic similarity.
Data extraction
The ingestion layer that recovers usable structure and provenance.
Memory
The persistent context layer beneath individual requests.
Alignment
The safety, policy, monitoring, and control layer around the complete system.
Explore the stack
Use a theme lane, search by project or idea, and change the ranking signal.
190 projects
Tools & agent workbenches
Runtimes, coding agents, orchestration frameworks, and operator surfaces where AI work actually happens.
openclaw
openclaw
Its rapid adoption shows demand for a persistent, cross-platform agent runtime that connects models, tools, channels, and personal workflows.
superpowers
obra
Reusable skills are becoming an engineering methodology: agents load bounded procedures instead of improvising every workflow from a prompt.
skills
mattpocock
Production-tested engineering procedures packaged as skills make expert workflows portable across agent clients and teams.
ECC
affaan-m
Skills, memory, hooks, and security checks package an operating method that can travel across coding-agent clients and repositories.
hermes-agent
NousResearch
Hermes makes learning from completed work part of the runtime, turning successful procedures into durable capabilities.
opencode
anomalyco
An open coding-agent client provides another inspectable execution surface for repository work and integration with the surrounding model stack.
dify
langgenius
Dify packages workflows, RAG, tools, and deployment into a shared product surface instead of leaving teams with isolated prompts.
langchain
langchain-ai
LangChain remains a central abstraction layer for models, tools, retrieval, and production agent plumbing.
claude-code
anthropics
Terminal-native coding agents turn codebase work into a tool-executing, reviewable loop instead of a chat transcript.
awesome-llm-apps
Shubhamsaboo
A large runnable example index shows which RAG, agent, and workflow patterns developers actually clone and remix.
gstack
garrytan
Opinionated specialist roles package product, design, engineering, release, and QA work into a repeatable agent operating model.
codex
openai
Codex makes local repository work goal-scoped, command-aware, and evidence-oriented.
browser-use
browser-use
Browser automation is becoming a first-class capability layer for agents that inspect and act on live web apps.
gemini-cli
google-gemini
A model-backed CLI with tool access and MCP support fits the shift from model choice to routable development harnesses.
OpenHands
OpenHands
A self-hosted workspace coordinates coding agents and automations across local and remote backends, making execution an operator-managed service.
deer-flow
bytedance
Sandboxes, memory, tools, skills, and subagents are converging into one harness for tasks that run for minutes or hours.
openai-cookbook
openai
Official examples turn API features into reusable patterns for evaluation, tool use, retrieval, and multimodal work.
ruflo
ruvnet
An execution layer around coding agents combines specialist roles, shared memory, plugins, and verification for coordinated repository work.
cline
cline
Cline's move across IDE, CLI, and SDK surfaces shows coding agents becoming embeddable infrastructure rather than one editor feature.
warp
warpdotdev
The terminal is becoming an operator cockpit where agents, commands, project context, and feedback loops meet.
autogen
microsoft
AutoGen helped normalize agentic programs as explicit conversations, tools, and coordination patterns.
crewAI
crewAIInc
Crew-style roles remain a practical entry point for decomposing agent work into specialist responsibilities.
goose
aaif-goose
Goose represents the open-source move from suggestions to installed, executing, tool-using desktop and code agents.
claude-cookbooks
anthropics
Cookbooks turn model capabilities into runnable patterns for tool use, retrieval, multimodal work, and production integration.
chrome-devtools-mcp
ChromeDevTools
Exposing Chrome inspection and debugging through MCP gives coding agents browser evidence for testing rendered behavior and diagnosing failures.
marketingskills
coreyhaines31
Reusable procedures for analytics, SEO, and conversion work turn audience-development tasks into bounded workflows an operator can review.
humanizer
blader
A reusable writing-review procedure makes recurring language problems explicit and gives newsroom agents a concrete revision checklist.
scientific-agent-skills
K-Dense-AI
Domain procedures and scientific database integrations show how research expertise can be packaged for reuse across agents that support the skills standard.
agent-browser
vercel-labs
A browser CLI designed around agent calls makes web interaction scriptable, inspectable, and easier to embed in coding and operations loops.
agno
agno-agi
Agno points toward full agent platforms with models, memory, tools, teams, and observability in one stack.
langgraph
langchain-ai
Graph-shaped state, checkpoints, and control flow are becoming the durable architecture behind long-running agents.
openhuman
tinyhumansai
Local-first personal memory plus fleet orchestration points toward assistants becoming durable operating environments rather than isolated chats.
awesome-copilot
github
Shared instructions, skills, and agent configurations show development practice moving from private prompting into reusable operational assets.
herdr
herdrdev
A terminal multiplexer with agent-aware states makes parallel coding sessions observable as work rather than anonymous panes.
claude-plugins-official
anthropics
Curated plugins turn domain workflows and integrations into installable capability packages with explicit ownership.
financial-services
anthropics
Reference agents, connectors, and skills for financial work show how professional teams can be packaged as inspectable domain capability modules.
prompt-optimizer
linshenkx
A dedicated prompt workbench makes revision and comparison reusable tasks instead of leaving prompt changes scattered across chat sessions.
oh-my-codex
Yeachan-Heo
Hooks, agent teams, and operator HUDs extend a coding agent into a configurable multi-agent workbench.
skills
vercel-labs
A cross-client installer and discovery layer signals that agent skills are becoming a package ecosystem rather than copied folders.
openai-agents-python
openai
The SDK packages tools, handoffs, tracing, and guardrails as code instead of one-off prompt structure.
deepagents
langchain-ai
DeepAgents packages planning, subagents, file state, and long-running behavior into a batteries-included harness.
mastra
mastra-ai
Mastra shows the JavaScript stack absorbing agents as production application components.
agenticSeek
Fosowl
A fully local browsing and coding agent reflects demand for private autonomy without per-task hosted API dependence.
symphony
openai
Isolated implementation runs move the operator from supervising every command to managing goals, evidence, and completed work.
skills
openai
An official skills catalog moves reusable Codex procedures into a discoverable, versionable capability layer.
VoiceStudio
debpalash
Local speech recognition, synthesis, dubbing, and an audio API offer a private production lane for transcripts and narrated editorial artifacts.
pydantic-ai
pydantic
Typed dependencies, structured outputs, tools, and eval integrations bring agent construction closer to ordinary production Python engineering.
skills
Provider-maintained procedures for products and technologies make operational knowledge directly loadable by agents.
open-seo
every-app
Search audits, rank tracking, and AI visibility exposed through MCP make the site's discovery loop accessible to agents as concrete tools.
camel
camel-ai
CAMEL keeps the research thread alive around scaling laws, roles, societies, and cooperation among agents.
rowboat
rowboatlabs
An open coworker that combines tools with persistent memory reflects the shift from chat assistants to accumulating work relationships.
ego-lite
citrolabs
Separate agent spaces inside a browser with existing login state explore how people and agents can work concurrently without competing for tabs.
eigent
eigent-ai
Open agent desktops are becoming operator workspaces where local tools, files, and multi-step execution meet.
InsForge
InsForge
A backend surface designed for agentic coding lets agents provision data, auth, storage, compute, and deployment through one explicit contract.
freebuff
CodebuffAI
Specialized agents across terminal, desktop, web, and repository workflows offer a case study in how agent access and distribution models evolve.
kimi-cli
MoonshotAI
Another capable terminal agent reinforces the trend toward model-swappable development harnesses with tools and repository access.
copilot-sdk
github
An SDK turns the coding agent from a destination product into a capability that other applications can embed.
feynman
advaitpaliwal
A packaged research agent with persistent sessions and a local workbench makes long-running investigation a dedicated operator workflow.
plugins
cursor
An official plugin specification makes coding-agent extensions installable and governed instead of hidden in personal configuration.
worktrunk
max-sixty
Agent-oriented worktree management makes isolated parallel coding sessions reproducible and easier to supervise.
agent-scripts
steipete
Small reusable scripts make recurring maintenance and review workflows callable by agents instead of rediscovered ad hoc.
plugins
openai
Public plugin manifests and examples show how skills, connectors, agents, and supporting assets become installable workflow packages.
agent-starter-pack
GoogleCloudPlatform
Agent templates with CI/CD, evaluation, and observability compress the path from demo to governed cloud deployment.
agents-cli
Creating, evaluating, and deploying agents through one CLI makes the lifecycle a reproducible engineering workflow.
maka
apache
An append-only runtime log makes tool calls, permission decisions, recovery, and evaluation evidence part of the agent workspace's core design.
mistral-vibe
mistralai
A minimal coding CLI from a model provider shows the harness layer becoming as strategic as the model endpoint.
skills-manager
xingkongliang
A central skill library with deployment and update tracking makes agent capabilities easier to inventory, synchronize, and maintain across clients.
agentos
rivet-dev
A library-level runtime using isolates lets applications host bounded agents inside existing backends without a separate SaaS control plane.
skills
BuilderIO
Portable planning and visual-recap skills make graphical artifacts part of the coding-agent coordination loop.
gemini-skills
google-gemini
Provider-maintained skills make model and SDK expertise loadable as procedures instead of copied documentation.
cookbook
cursor
Official workflow recipes expose the operating patterns around coding agents, not only the editor feature set.
claude-plugins-community
anthropics
A reviewed community marketplace shows how agent capabilities can be distributed with explicit manifests and a controlled submission pipeline.
openai-agents-js
openai
A JavaScript SDK brings tools, handoffs, tracing, and voice agents into the dominant web application runtime.
agentfield
Agent-Field
Treating agents as routable APIs and microservices puts memory, asynchronous execution, scaling, and audit into one control plane.
agentkit
coinbase
Wallet primitives let agents transact under explicit tool contracts, pushing identity, permissions, and audit into the agent economy.
Context engineering & memory
The supply chain that turns documents, code, web pages, tools, and knowledge graphs into usable agent context.
markitdown
microsoft
File-to-Markdown conversion is becoming the dependable ingestion layer agents need before retrieval, synthesis, and audit trails.
firecrawl
firecrawl
Web extraction is moving from scraping glue to an agent-ready primitive for monitoring, RAG, and newsroom intake.
awesome-mcp-servers
punkpeye
MCP server catalogs show how quickly agent capabilities are becoming installable infrastructure.
servers
modelcontextprotocol
The reference server collection makes tools and context sources portable across agent clients.
rtk
rtk-ai
Filtering command output before it enters model context makes terminal verbosity a controllable cost in coding-agent loops.
Agent-Reach
Panniantong
One CLI for difficult web and social sources makes external evidence retrieval a reusable agent capability rather than per-source scraping glue.
codegraph
colbymchenry
A continuously synchronized local code graph reduces repeated repository scans and gives coding agents durable structural context.
docling
docling-project
Production context pipelines need structure-preserving conversion for PDFs and office documents before retrieval can be trustworthy.
last30days-skill
mvanhorn
A bounded skill for recent cross-platform research packages source discovery, recency, and synthesis into a repeatable evidence workflow.
GitNexus
abhigyanpatwari
A browser-local repository graph combines privacy, structural navigation, and Graph RAG without requiring a hosted indexing service.
langextract
Structured extraction with precise source grounding connects probabilistic models to auditable data pipelines.
OpenViking
volcengine
A virtual filesystem unifies resources, memories, and skills, letting agents browse summaries and load only the context a task needs.
cli
googleworkspace
Discovery-driven commands, JSON output, and skills expose office workflows to agents; the repository explicitly does not promise official product support.
graphiti
getzep
Temporal knowledge graphs make memory updateable, queryable, and less tied to one chat window.
cognee
topoteretes
Cognee turns memory into a self-hosted knowledge graph layer, matching New Runtime's shared-context thesis.
supermemory
supermemoryai
A local-capable memory API separates persistent context from any one model, client, or conversation.
opendataloader-pdf
opendataloader-project
AI-ready PDF parsing is becoming a dedicated infrastructure layer because document layout and provenance cannot be recovered after naive text extraction.
agentmemory
rohitg00
Shared persistent memory across coding agents reduces repeated onboarding and makes context continuity measurable against real tasks.
repomix
yamadashy
Packing repositories into AI-readable context remains a practical bridge between large codebases and coding agents.
fastmcp
PrefectHQ
FastMCP lowers the cost of turning internal APIs and scripts into agent-callable tools.
haystack
deepset-ai
Haystack keeps context engineering explicit through pipelines, retrieval, routing, memory, and observability.
Unlimited-OCR
baidu
One-shot long-horizon document parsing challenges page-by-page pipelines and reduces the reconstruction burden in document context systems.
WeKnora
Tencent
Document retrieval, reasoning agents, and a versioned wiki connect one-off answers to a knowledge base that can be edited and maintained over time.
context-mode
mksglu
Sandboxing tool output and routing context through MCP and hooks turns token pressure into an explicit systems optimization problem.
olmocr
allenai
High-quality PDF linearization improves the training and retrieval substrate before an LLM ever sees the document.
turbovec
RyanCodrai
A quantized vector index with online ingest and incremental persistence addresses the memory and maintenance cost of growing retrieval corpora.
Memori
MemoriLabs
Agent-native memory that writes structured state into existing data infrastructure matches the move from chat history to governed organizational memory.
openwiki
langchain-ai
Agent-maintained repository documentation creates a durable context layer that stays closer to the changing code.
memU
NevaMind-AI
Memory shared across agents addresses the fragmentation caused when every client maintains a separate understanding of the user.
LEANN
StarTrail-org
Storage-efficient local retrieval makes private personal and edge-device RAG practical without a heavyweight vector database.
semantica
semantica-agi
Knowledge graphs, ontology management, and decision provenance connect document ingestion to traceable reasoning and analytical memory.
claude-context
zilliztech
Repository-wide semantic search exposed through MCP gives multiple coding clients a shared context service.
cocoindex
cocoindex-io
Incremental indexing keeps long-running agent context fresh without paying for full reprocessing after every source change.
git-mcp
idosal
A remote MCP context layer gives agents source-grounded access to repositories without manually packing every codebase into the prompt.
pxpipe
teamchong
Rendering bulky text context as images explores a new tradeoff between multimodal input capacity, token cost, and recoverability.
apify-mcp-server
apify
A shared MCP surface makes existing crawlers and extraction tools callable by agents, with structured results that can feed evidence intake.
airweave
airweave-ai
A dedicated retrieval layer separates continuously changing source synchronization from the agent application itself.
engram
Gentleman-Programming
SQLite-backed memory exposed through CLI, HTTP, and MCP gives different coding agents a shared, inspectable store beyond a single session.
semble
MinishLab
Token-efficient structural code search reduces the context tax of finding the right repository evidence before editing.
mgrep
mixedbread-ai
Semantic grep across code, images, and documents offers agents one retrieval primitive across heterogeneous project context.
Acontext
memodb-io
Treating skills as memory connects reusable procedures with the state and evidence agents need to apply them correctly.
hyperresearch
jordan-gibbs
A persistent research vault, citation checks, and source-independence review connect fresh research to reusable evidence and accountable synthesis.
seekdb
oceanbase
Unifying vector, text, structured, and semi-structured retrieval reduces the storage fragmentation behind agent memory.
files-to-prompt
simonw
A small deterministic context packer remains useful because inspectable preprocessing is often preferable to opaque ingestion.
chunkhound
chunkhound
Local-first codebase intelligence gives agents structural retrieval without exporting private repositories to a hosted index.
work-iq
microsoft
An MCP and CLI surface for organizational work context points toward governed enterprise knowledge becoming directly agent-callable.
LLM-UI & generated interfaces
Interfaces assembled around model output, task state, validation, and human correction.
screenshot-to-code
abi
Turning screenshots and recordings into editable frontend code makes visual references a direct input to interface-building workflows.
impeccable
pbakaus
An explicit design language helps coding agents reason about visual tradeoffs instead of reproducing generic component patterns.
archify
tt-a1i
Typed diagram data compiled into HTML and SVG makes architecture explanations inspectable, diffable, and reusable by agents and editors.
pretext
chenglou
Fast deterministic text measurement gives generative interfaces a layout primitive agents can calculate and render instead of approximating.
hyperframes
heygen-com
HTML-to-video gives agents a code-native visual medium they can generate, inspect, revise, and render deterministically.
ToolJet
ToolJet
Agents generate pages and queries against component and data contracts, leaving an operational app that people can continue editing visually.
diagram-design
cathrynlavery
Reusable HTML and SVG diagram procedures give research and engineering agents a concrete way to explain systems, decisions, and feedback loops.
OpenMAIC
THU-MAIC
An interactive classroom turns source material into coordinated teaching agents, making adaptive learning a useful case study for generated interfaces.
design.md
google-labs-code
A portable design-system file gives coding agents durable visual context instead of forcing them to infer taste from screenshots on every task.
json-render
vercel-labs
Structured UI output lets model answers become inspectable interfaces instead of unstructured prose blobs.
tambo
tambo-ai
A React generative-UI SDK points to applications where the interface is assembled around the task state.
magentic-ui
microsoft
Browser and local-file agents need an operator interface, not just a backend chain of tool calls.
stitch-skills
google-labs-code
Design procedures packaged as portable skills connect generative UI tools to repeatable production workflows.
open-slide
1weiho
A code-native slide framework gives agents structured layout, components, and renderable output rather than opaque presentation files.
drawio-mcp
jgraph
Streaming diagram primitives through MCP turns model-generated structure into an inspectable visual work surface while it is being built.
openai-apps-sdk-examples
openai
Runnable Apps SDK examples show how model responses become interactive, stateful product surfaces instead of plain text.
realtime-voice-component
openai
A reusable voice component makes low-latency multimodal interaction an embeddable interface primitive rather than a bespoke demo.
generative-ui
CopilotKit
Examples across AG-UI, A2UI, Open JSON UI, and MCP Apps reveal an emerging protocol layer for model-generated interfaces.
taste-skill
senlindesign
Extracting design decisions and tradeoffs rather than only tokens gives interface agents a more durable representation of visual intent.
Evals, specs & reliability
The control plane: specifications, tests, protocols, structured output, security boundaries, and sandboxes.
andrej-karpathy-skills
multica-ai
A compact CLAUDE.md derived from recurring coding-agent failures shows how operational lessons are becoming reusable behavioral contracts.
system-prompts-and-models-of-ai-tools
x1xhlol
The collection exposes how production AI tools combine prompts, tool contracts, and hidden orchestration, making harness design inspectable.
ponytail
DietrichGebert
Encoding senior-engineer restraint into agent behavior addresses the growing cost of unnecessary abstractions and generated-code sprawl.
spec-kit
github
Spec-driven development makes requirements, plans, and acceptance checks the control surface before agents write code.
autoresearch
karpathy
A bounded single-GPU research loop turns hypothesis, experiment, measurement, and iteration into an inspectable autonomous workflow.
i-have-adhd
ayghri
Explicit output rules make agent responses easier to act on, treating answer structure and cognitive load as part of the interaction contract.
A2A
a2aproject
Agent-to-agent protocols matter as soon as agents become opaque services owned by different products or teams.
agentskills
agentskills
A shared skill specification makes procedures portable across agent clients and gives teams a stable contract for capability packaging.
evals
openai
Evals are moving from model benchmarks into product infrastructure for judging complete AI systems.
deepeval
confident-ai
DeepEval makes LLM and agent quality measurable in CI-style loops instead of relying on manual prompt taste.
agent-lightning
microsoft
Agent training is shifting from prompt tweaking toward trace-level optimization of complete trajectories and tool decisions.
SkillSpector
NVIDIA
As skills become executable supply-chain artifacts, static inspection for malicious instructions and risky behavior becomes mandatory.
outlines
dottxt-ai
Grammar and schema-constrained outputs reduce the gap between probabilistic generation and software contracts.
CubeSandbox
TencentCloud
Isolated execution environments are required before agents can safely run code, browsers, or external tools.
hexstrike-ai
0x4m4
Offensive MCP tooling signals that agent tool access is both a capability frontier and a security boundary.
improve
shadcn
Separating expensive architectural review from cheaper execution models turns model routing into a concrete engineering workflow.
monty
pydantic
A minimal Rust-based Python interpreter gives agents a deliberately bounded execution surface instead of a full host runtime.
evidently
evidentlyai
Evaluation, testing, and monitoring are converging into one operational loop for AI systems after deployment.
AI-Infra-Guard
Tencent
Scanning agents, skills, MCP servers, and model interfaces brings several AI attack surfaces into one inspectable assessment workflow.
sandbox
agent-infra
Combining browser, shell, files, MCP, and an editor in one container gives agent runs a bounded and reproducible execution environment.
zerolang
vercel-labs
A language built around structured diagnostics gives coding agents machine-readable repair signals instead of prose compiler output.
aidlc-workflows
awslabs
A harness-neutral development workflow packages planning and verification rules so delivery practice can remain consistent across coding agents.
skills
humanlayer
Procedures for instruction repair and iterated coding workflows make sensors, actions, and verification explicit parts of repository automation.
skillsbench
benchflow-ai
Skills need behavioral benchmarks that measure both the procedure and the agent's ability to invoke it correctly.
skills
AI-Builder-Club
Codebase maps and loop-engineering procedures package structural evidence and repeatable maintenance cycles for coding agents.
openevals
langchain-ai
Ready-made evaluators lower the cost of putting repeatable quality checks around LLM and agent behavior.
ClawKeeper
SafeAI-Lab-X
Skills, plugins, and runtime watchers form a layered security boundary around a tool-using personal agent.
agent-protocol
langchain-ai
A framework-neutral service contract separates agent infrastructure and lifecycle operations from any one orchestration library.
AutomationBench
zapier
Realistic business automation tasks test whether agents can complete multi-step work across tools, not merely answer benchmark questions.
MemoryData
OpenDataBox
A unified benchmark for memory-augmented agents makes storage, extraction, retrieval, and use comparable as a system.
fully-automated-prompt-optimization
cisco-foundation-ai
Optimizing multi-step LLM chains with coding agents and evals moves improvement from manual prompt editing to a measured loop.
altk-evolve
AgentToolkit
Iteration-based self-improvement makes agent changes explicit artifacts that can be compared and evaluated between runs.
Model & serving infrastructure
Serving throughput, local training, cache economics, and the compute layer underneath agent systems.
ollama
ollama
A local model runner and client libraries provide a practical private inference lane for tools, experiments, and personal agents.
vllm
vllm-project
High-throughput serving keeps model routing and self-hosted inference economically plausible.
unsloth
unslothai
Faster local training and inference give teams a cheaper private lane for model experiments.
minimind
jingyaogong
A compact PyTorch training stack exposes pretraining, fine-tuning, alignment, and tool-use work as reproducible experiments on smaller models.
llmfit
AlexsJones
Hardware inspection and measured local benchmarks help choose models that fit available memory and throughput before committing to a deployment.
browser
lightpanda-io
A headless browser built for automation treats browser execution cost and throughput as first-class agent infrastructure.
modular
modular
MAX serving, model pipelines, kernels, and Mojo connect deployment decisions to the lower-level compute stack behind model applications.
omlx
jundot
Continuous batching and memory-to-SSD context caching address the repeated prompts and long sessions of local coding-agent workloads.
web-llm
mlc-ai
WebGPU inference brings model execution into the browser, enabling private client-side interfaces without a separate inference server.
LMCache
LMCache
KV-cache infrastructure turns long context and repeated workloads into a serving optimization problem.
bifrost
maximhq
A shared API with failover, load balancing, and caching centralizes the operational controls needed when several agents share model providers.
semantic-router
vllm-project
Request signals, user preferences, and policy guide model paths across heterogeneous inference infrastructure, separating routing decisions from application code.
experiential
experientiallabs
A shared gateway combines model access, identity and budget controls, and traffic-driven routing experiments for agent workloads.
magnitude
magnitudedev
Hardware profiling, model recommendations, and local serving connect an existing coding harness to models suited to the user's machine.
router
weave-os
An on-box scorer routes individual model requests across providers, making model selection a policy layer that can be evaluated independently of the agent.
marin
marin-community
Recorded data preparation, training, evaluation, and experiments make the model-building process available for inspection alongside its outputs.
Switchyard
NVIDIA-NeMo
Embeddable routing algorithms and gateway plugins let teams test model selection while keeping their existing transport and credential handling.
MTPLX
youssofal
Using a model's own multi-token prediction heads explores faster local inference without reserving memory for a separate draft model.
What has to reproduce
The JSON route exposes the ranked repository dataset, proof contract, handoff routes, and public GitHub source URLs.
- Verification
- locally reproduced
- Outputs
- 6 fixtures
- Replay
- 2 receipts
- Axes
- 7 assessed
Fixture outputs
- Structured radar dataset
dist/projects/github-projects-for-ai-engineers.jsonThe JSON route exposes the ranked repository dataset, proof contract, handoff routes, and public GitHub source URLs.
Open route - Human radar page
dist/projects/github-projects-for-ai-engineers/index.htmlThe HTML radar renders the same project count, theme lanes, and machine-readable handoff links as the dataset.
Open route - Agent build brief
dist/projects/github-projects-for-ai-engineers/build-brief.mdThe build brief names the data source, GitHub metric refresh boundary, privacy constraints, and acceptance checks.
Open route - Clean-room replay fixture
scripts/project-proof-depth-v2-fixture-test.mjsThe fixture replays the bounded deterministic-ranking task in a temporary directory, proves the initial failure of the wrong sort order, proves the final pass of the stars-then-forks rule, and verifies that an unrelated file is unchanged.
- Independent replay kit
scripts/project-proof-independent-replay-kit.mjsThe kit starts with a ranking function that violates the published sort rule, exposes no ready solution, protects every out-of-scope file by checksum, and refuses to emit a receipt without a non-author attestation and a passing final check.
- Independent agent replay receipt
src/data/project-proof-receipts/github-projects-for-ai-engineers-independent-agent-v1.jsonA separately spawned agent received only the isolated kit, observed the failing deterministic-ranking check, changed only src/rank.mjs, passed the final check, and produced an attestation that the primary verifier accepted.
Open route
Replay commands
npm run project-proof:replay:testnpm run project-proof:replay:prepare -- --project=github-projects-for-ai-engineersnpm run validate:contentnpm run buildnpm run validate:agent-indexes
Provider scopes
- GitHubread
Public repository metadata and repository pages only.
Stars, forks, archived state, redirects, and canonical full names drift over time.
Human approves a live GitHub refresh before network calls.
- Local Telegram exportread
Repository-link roots and duplicate counts from an approved local export snapshot.
The radar starts from private discovery provenance but publishes only public GitHub repository URLs.
Human names the export snapshot or existing dataset before extraction.
- Public sitepublish
Project content, project dataset, and generated static routes for this radar only.
The public radar must expose the refreshed structured dataset and human visualization together.
Owner approves commit, push, and deploy separately from the metric refresh.
Seven-axis Buildability
non author replayed · 2026-08-15
- timehigh
- The clean-room ranking task reaches a checked result in one bounded local run, while the public blueprint targets one local content pass plus a GitHub metric refresh.
- BottleneckA future radar refresh still needs fresh GitHub metrics for every retained repository.
- code burdenmedium
- The replay needs one focused comparator edit plus the existing test, and the public blueprint needs normalization scripts and content validation rather than a bespoke application.
- BottleneckCurating theme lanes and relevance still needs an editor who knows the AI-engineering landscape.
- integration burdenmedium
- The replay uses only repository files, Node, and a test runner, while the wider blueprint touches public GitHub reads and Astro content routes that already exist in this repository.
- BottleneckGitHub metric drift forces every refresh through an approved external read before the ledger is current.
- operational burdenhigh
- The first proof is local and temporary, requires no server, and touches no shared or production state.
- BottleneckKeeping the published radar current requires a recurring refresh protocol rather than one-off runs.
- permission clarityhigh
- GitHub read, local Telegram-export read, and public-site publish scopes are named separately with explicit approval gates in the build brief and acceptance contract.
- BottleneckA live metric refresh and a public deploy still need two separate owner approvals.
- reproducibilityhigh
- A separately spawned agent received only the isolated public kit, reproduced the failing deterministic-ranking check, changed the one allowed comparator file, and passed the same acceptance check without inspecting the author solution.
- BottleneckThe bounded fixture proves the published sort rule, not a full 140-repository refresh against live GitHub data.
- failure recoveryhigh
- The replay records the initial failing check, preserves an unrelated file byte-for-byte, performs no remote action, and leaves explicit residual-risk and approval sections.
- BottleneckRolling back a bad public radar release is outside this local proof.
Replay evidence
Bounded receipts, with limitations kept visible.
automated clean room · 2026-08-15
npm run project-proof:replay:test- This replay proves the deterministic ranking contract on a bounded fixture dataset, not a real external repository integration or a live GitHub metric refresh.
- No human or separately operated agent independently interpreted the brief in this clean-room replay.
independent agent · 2026-08-15
node --test test/rank.test.mjs- This proves a bounded non-author replay of the public workflow contract, not integration into a production repository.
- Reviewer identity is a local agent attestation and is not cryptographically verified.
- GitHub stars and forks are time-sensitive; the public method must keep the refresh date visible.
- The first pass does not clone, run, or security-audit the listed repositories.
- Export-hit counts are discovery signals, not public provenance; raw Telegram data remains private.
- The ecosystem map is a navigation and architecture aid, not a compatibility matrix or recommendation to adopt every tool.
- The bounded replay proves the deterministic ranking rule on a small fixture dataset, not a full 140-repository refresh against live GitHub metrics.
- The independent replay proves the bounded public workflow contract, not integration into a real external or production repository.
- The independent reviewer identity is a local agent attestation and is not cryptographically verified.
Weekly knowledge desk
Weekly updates
2 reviewed updates to this project, newest first. Every update is checked against its evidence before it changes the record.
LingBot-Map and ghidra-mcp add inspectable AI projects to the radar
Last week’s items described LingBot-Map with model access, run instructions, and evaluation scripts, and identified ghidra-mcp as a public repository for a Ghidra MCP server. These concrete repository examples fit the guide’s focus on mapping projects and the wider production AI stack.
- Robbyant’s LingBot-Map repository presents a feed-forward model for streaming 3D reconstruction and provides signal, Oct 9, 2026
- ghidra-mcp: Ghidra MCP Server signal, Oct 8, 2026
Docker and Cloudflare published AI-engineering projects
Docker published a public repository of reusable skills for AI coding agents, and Cloudflare open-sourced Cloudflare OS, an AI productivity environment. These are concrete repository observations in the AI-engineering ecosystem the guide aims to track and map.
- skills: A collection of Docker skills for AI coding agents to help them build, test, debug, and optimize… signal, Oct 1, 2026
- cloudflare/cloudflare-os signal, Oct 3, 2026
Retrieval answer
A source-backed GitHub growth radar plus a map of the models, orchestration, retrieval, memory, security, automation, and infrastructure layers around production AI systems. Turn public repository observations and ecosystem references into a ranked repository radar and a navigable map of the wider production AI stack.