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★.
| Project | Stars | In one line |
|---|---|---|
| deepseek-ai/deepseek-harness | 215.4k | “everything is a plugin” |
| anomalyco/opencode | 205.8k | open-source terminal agent |
| anthropics/claude-code | 144.4k | fastest shipper in the group |
| google-gemini/gemini-cli | 106.9k | Google’s agent CLI |
| OpenHands/OpenHands | 86.8k | autonomous dev agent |
| cline/cline | 67.7k | extension → own desktop app |
| aaif-goose/goose | 54.0k | Block’s desktop agent |
| Aider-AI/aider | 48.8k | pioneer, now stalling |
| charmbracelet/crush | 28.0k | TUI + nightly channel |
| xai-org/grok-build | 26.6k | embeds into editors via ACP |
| anywhere-labs/dsh-desktop | 24.3k | desktop for the DSH ecosystem |
| HKUDS/DeepCode | 16.5k | harness + 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.
| Project | Stars | In one line |
|---|---|---|
| langchain-ai/langchain | 145.9k | the platform framework |
| microsoft/autogen | 60.9k | Microsoft’s multi-agent |
| FoundationAgents/MetaGPT | 70.3k | “AI software company” |
| crewAIInc/crewAI | 58.2k | role-based crews |
| HKUDS/nanobot | 47.9k | minimal, self-hosted |
| openai/openai-agents-python | 29.3k | OpenAI’s official SDK |
| pydantic/pydantic-ai | 19.8k | type-safe agents |
| camel-ai/camel | 17.7k | research lineage |
| QwenLM/Qwen-Agent | 17.1k | the 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.
| Project | Stars | In one line |
|---|---|---|
| punkpeye/awesome-mcp-servers | 94.6k | the server catalog |
| modelcontextprotocol/servers | 90.2k | official servers |
| headroomlabs-ai/headroom | 70.4k | MCP infrastructure |
| DeusData/codebase-memory-mcp | 42.6k | codebase as memory |
| microsoft/mcp-for-beginners | 17.2k | the curriculum |
| hangwin/mcp-chrome | 12.4k | browser as server |
| modelcontextprotocol/registry | 7.2k | the 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.
| Project | Stars | In one line |
|---|---|---|
| zed-industries/zed | 89.9k | Rust, ACP birthplace |
| neovim/neovim | 102.2k | the agent-plugin base |
| coder/code-server | 79.2k | VS Code on a server |
| continuedev/continue | 35.8k | extension → agent |
| helix-editor/helix | 46.1k | modal, minimal |
| lapce/lapce | 38.8k | Rust |
| TabbyML/tabby | 33.9k | self-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.
| Project | Stars | In one line |
|---|---|---|
| ollama/ollama | 180.4k | easiest local models |
| ggml-org/llama.cpp | 127.5k | the original runtime |
| vllm-project/vllm | 91.2k | high-performance serving |
| exo-explore/exo | 47.3k | personal-device clusters |
| janhq/jan | 44.4k | desktop 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.
| Project | Stars | In one line |
|---|---|---|
| junegunn/fzf | 82.9k | fuzzy finder |
| BurntSushi/ripgrep | 68.1k | fast grep |
| sharkdp/bat | 60.4k | modern cat |
| starship/starship | 59.8k | polyglot prompt |
| sxyazi/yazi | 42.0k | TUI file manager |
| ast-grep/ast-grep | 15.8k | AST-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.
| Project | Stars | In one line |
|---|---|---|
| guillaumemeyer/watermarks-remover | 21.3k | privacy-first |
| CopilotKit/OpenBot | 4.4k | AI coworkers with their own machine |
| Hisn00w/ASu-skills | 4.0k | job-specific skills |
| yetone/cumora | 3.5k | chat for agent teams |
| Leonxlnx/unlazy | 3.2k | discipline-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.txtis 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.