~/blog/agentic-landscape-50-projects
tech notes #landscape#agents#oss#mcp#workflow Wakii team 7 min read

50 projects shaping agentic coding — the September 2026 map

50 projects shaping agentic coding — the September 2026 map

A year ago, “agentic coding” was a question. Today it is an industry with its own map. We swept the GitHub API on 2026-09-08, filtered fifty projects into seven groups, and wrote the picture down here — with star counts exactly as retrieved, because these numbers move weekly.

TL;DR: the harness layer is racing hardest but nobody has broken away; MCP has become the default integration layer; “skills” are turning into a new distribution unit; local inference is no longer a fringe; terminals and editors are swapping roles.

Harnesses and coding agents (12)

The most competitive layer. DeepSeek Harness — a newcomer in August with an “everything is a plugin” manifesto — already sits at 215.4k★. opencode leads at 205.8k★ after roughly sixteen months. claude-code holds 144.4k★ with a three-releases-in-four-days cadence. The counterpoint: aider, once the icon of the category, has been nearly frozen for a year at 48.8k★.

ProjectStarsIn one line
deepseek-ai/deepseek-harness215.4k“everything is a plugin”
anomalyco/opencode205.8kopen-source terminal agent
anthropics/claude-code144.4kfastest shipper in the group
google-gemini/gemini-cli106.9kGoogle’s agent CLI
OpenHands/OpenHands86.8kautonomous dev agent
cline/cline67.7kextension → own desktop app
aaif-goose/goose54.0kBlock’s desktop agent
Aider-AI/aider48.8kpioneer, now stalling
charmbracelet/crush28.0kTUI + nightly channel
xai-org/grok-build26.6kembeds into editors via ACP
anywhere-labs/dsh-desktop24.3kdesktop for the DSH ecosystem
HKUDS/DeepCode16.5kharness + multi-agent loop

Multi-agent frameworks (9)

The layer beneath harnesses. Model-attached frameworks (Qwen-Agent) and minimalists (nanobot) are taking seats from the previous research generation.

ProjectStarsIn one line
langchain-ai/langchain145.9kthe platform framework
microsoft/autogen60.9kMicrosoft’s multi-agent
FoundationAgents/MetaGPT70.3k“AI software company”
crewAIInc/crewAI58.2krole-based crews
HKUDS/nanobot47.9kminimal, self-hosted
openai/openai-agents-python29.3kOpenAI’s official SDK
pydantic/pydantic-ai19.8ktype-safe agents
camel-ai/camel17.7kresearch lineage
QwenLM/Qwen-Agent17.1kthe Qwen ecosystem

MCP — the new integration layer (7)

MCP is replaying the role HTTP once played for the web: the best-known server list sits at 94.6k★, the official registry at 7.2k★, Microsoft wrote a curriculum. Both directions are growing — Chrome becomes a server, codebases become memory.

ProjectStarsIn one line
punkpeye/awesome-mcp-servers94.6kthe server catalog
modelcontextprotocol/servers90.2kofficial servers
headroomlabs-ai/headroom70.4kMCP infrastructure
DeusData/codebase-memory-mcp42.6kcodebase as memory
microsoft/mcp-for-beginners17.2kthe curriculum
hangwin/mcp-chrome12.4kbrowser as server
modelcontextprotocol/registry7.2kthe registry

Editors and IDEs (7)

zed alone at 89.9k★ proves a from-scratch Rust editor has a seat. continue re-branded itself as an “open-source coding agent” (35.8k★) — the industry’s move in miniature: an editor is no longer where you type code, it is where you dispatch agents.

ProjectStarsIn one line
zed-industries/zed89.9kRust, ACP birthplace
neovim/neovim102.2kthe agent-plugin base
coder/code-server79.2kVS Code on a server
continuedev/continue35.8kextension → agent
helix-editor/helix46.1kmodal, minimal
lapce/lapce38.8kRust
TabbyML/tabby33.9kself-hosted completion

Local inference (5)

The numbers argue before anyone does: agentic work does not have to cross the cloud, and “my data stays on my machine” is a real pull.

ProjectStarsIn one line
ollama/ollama180.4keasiest local models
ggml-org/llama.cpp127.5kthe original runtime
vllm-project/vllm91.2khigh-performance serving
exo-explore/exo47.3kpersonal-device clusters
janhq/jan44.4kdesktop local AI

Terminal tooling (6)

This layer explains why TUI agents win so easily: the terminal stack is already there and very healthy. ast-grep — structural grep — is what several agent CLIs use internally.

ProjectStarsIn one line
junegunn/fzf82.9kfuzzy finder
BurntSushi/ripgrep68.1kfast grep
sharkdp/bat60.4kmodern cat
starship/starship59.8kpolyglot prompt
sxyazi/yazi42.0kTUI file manager
ast-grep/ast-grep15.8kAST-aware grep

The one-month-old wave (5 — created after Aug 8)

Most interesting to anyone building agents: unlazy (3.2k★) packages “completion discipline” — an acceptance ledger written first, runnable gates, reports backed by evidence — into a single installable skill (npx skills add, via vercel-labs’ skills CLI). The philosophy of “gates, not self-reports” is being boxed and shipped like a package.

ProjectStarsIn one line
guillaumemeyer/watermarks-remover21.3kprivacy-first
CopilotKit/OpenBot4.4kAI coworkers with their own machine
Hisn00w/ASu-skills4.0kjob-specific skills
yetone/cumora3.5kchat for agent teams
Leonxlnx/unlazy3.2kdiscipline-as-a-skill

Three patterns across the map

MCP has won the integration layer — no other protocol has a 94.6k★ catalog or a Microsoft curriculum; every tool that wants to be callable is shipping itself as an MCP server.

Skills are the new packaging unit — from job-hunting skill packs to an “anti-laziness” skill, a small human-readable format is being distributed like packages. Whoever owns a skill kit (like the nine-agent team in Wakii’s agents & kit) is standing on the fastest-growing layer.

Release cadence mirrors process — claude-code’s three releases in four days and aider’s year-long freeze are two ends of one axis; the decision lives in the pipeline, not the commit count.

Questions the map leaves open

  • opencode reached 205.8k★ in ~16 months with 26.8k forks — where does that velocity come from, and what keeps it after the novelty fades?
  • llms.txt is starting to appear at repo roots — is “AI-friendly repo” becoming a documentation standard?
  • Skills CLIs promise installs into “every detected agent” — who becomes the npm registry of skills, and where does quality control live?
  • aider has been frozen for a year yet keeps 48.8k★ — is that “finished” or “lost momentum”?
  • MCP took the integration layer, ACP targets the editor layer — which standard covers agents orchestrating each other, the layer where Wakii’s brackets, gates, and watchdog live?

This map is a snapshot from 2026-09-08 — it will drift from reality within weeks. But its shape — harnesses racing, MCP consolidating integration, skills becoming the packaging unit, local inference big enough to be a foundation — looks like the shape of at least the rest of 2026.

Wakii — an agentic IDE with a nine-agent team — joins this map from the editor layer, and builds itself with the very process the harness layer is standardizing. Grab a build at the download page.

Wakii is an agentic IDE with a built-in superpowers team.

get wakii ← all posts