| Dify |
136.3k |
Python, TypeScript |
workflow-DAG (graphon engine) |
SSRF proxy + code sandbox (2 layers) |
-- |
Docker (7+ containers) |
plugins + workflow nodes |
graphon engine, 7+ containers, 30+ vector DBs, plugin daemon |
workflow-engine |
production |
docker-compose |
1,283,000 |
Modified Apache 2.0 |
Layer-based graph engine middleware (graphon); LLM-powered metadata filtering; Child engine spawning for iteration |
| Claude Code |
109.6k |
TypeScript |
streaming tool pipeline |
per-tool permission + seatbelt sandbox |
4-layer cascade (lossless before lossy) |
seatbelt (macOS) |
MCP |
4-layer context cascade, hidden pet system (18 species) |
coding-agent |
production |
npm |
510,000 |
Proprietary |
4-layer context cascade: lossless before lossy; Streaming tool execution with RWLock; buildTool() factory over inheritance |
| OpenHands |
70.9k |
Python, TypeScript |
event-driven (observation/action) |
GraySwan + Invariant + LLM-analyzer (3 layers) |
10-condenser pipeline |
Docker |
-- |
10 condensers, 3-layer security, 487-line stuck detector |
coding-agent |
production |
docker |
400,000 |
MIT |
10-condenser pipeline: most sophisticated context management; 3-layer security: defense-in-depth; 487-line stuck detector for breaking loops |
| Cline |
60k |
TypeScript |
unified loop (3,756-line core) |
human-in-the-loop (1 layer) |
truncate |
-- |
MCP + hooks |
3,756-line core, 43 providers, YOLO mode, hooks system |
coding-agent |
production |
npm |
560,000 |
Apache-2.0 |
43 provider adapters with factory pattern; Hooks system for TaskStart/TaskResume/PreCompact events; Focus Chain task tracker |
| DeerFlow |
58.1k |
Python, TypeScript |
middleware pipeline (14 layers) |
-- |
summarize (hash loop kill@5) |
Docker |
-- |
14-layer middleware, hash-based loop detection |
workflow-engine |
beta |
docker |
-- |
MIT |
Middleware-first architecture: 14+ composable middlewares; Hash-based loop detection (warn@3, kill@5); DanglingToolCallMiddleware for orphan calls |
| MiroFish |
50.2k |
Python, Vue |
ReACT loop (OASIS/camel-ai) |
-- |
-- |
-- |
-- |
LLM social sim as 'collective intelligence', core is OASIS |
social-simulation |
beta |
pip |
38,800 |
AGPL-3.0 |
Dual-platform parallel simulation with independent action logs; Forced multi-tool ReACT strategy with unused tool reminders |
| Goose |
37.3k |
Rust, TypeScript |
MCP-first extension bus |
5-inspector pipeline (security, egress, adversary, permission, repetition) |
summarize (tool-pair background) |
process isolation |
MCP + 6 extension types |
MCP-first bus, 5-inspector pipeline, 30+ LLMs |
coding-agent |
production |
binary |
198,000 |
Apache-2.0 |
Declarative provider pattern via 10-line JSON; 31-key env var blocklist; Tool-pair background summarization |
| Pi Mono |
32k |
TypeScript |
game-engine loop (tick-based) |
-- |
summarize |
-- |
plugins (tool definitions) |
Game-engine arch, stealth mode as Claude Code |
coding-agent |
beta |
npm |
147,444 |
MIT |
Lazy provider loading via dynamic import + promise caching; Steering/follow-up two-lane input queue; TUI-integrated tool rendering |
| MemPalace |
30.1k |
Python |
layered recall (4-layer stack) |
-- |
4-layer cascade (~600 token wake-up) |
-- |
MCP |
4-layer memory stack, ChromaDB palace metaphor, +34% retrieval |
memory-system |
experimental |
pip |
15,534 |
MIT |
Layered loading pattern (~600 token wake-up); Auto-teach via MCP status tool response; PreCompact hook strategy; Temporal knowledge graph in SQLite |
| Lightpanda |
27.3k |
Zig, Rust |
comptime bridge (Zig-V8 zero-cost FFI) |
-- |
none (browser, not agent) |
-- |
MCP + CDP |
From-scratch browser for AI agents, 9x Chrome, 16x less memory |
browser |
beta |
binary |
91,000 |
AGPL-3.0 |
Comptime V8 bridge: zero runtime dispatch overhead; Arena-based memory (16x reduction); Integer bitcast CDP dispatch |
| OpenAI Codex CLI |
27k |
Rust |
queue-pair (Sender/Receiver decoupled) |
seatbelt + landlock + restricted-token + Guardian AI (4 layers) |
summarize (two-phase extraction) |
seatbelt + landlock + restricted token (3-OS) |
MCP + plugins |
Queue-pair arch, Guardian AI approval, 3-OS sandbox (17K LoC) |
coding-agent |
production |
binary |
549,000 |
Apache-2.0 |
Guardian AI risk-scoring with auto-approve threshold; Queue-pair decoupling core from frontends; Progressive sandbox escalation; Two-phase memory extraction |
| Hermes Agent |
26.1k |
Python |
single-file loop (9,000-line agent) |
memory threat detection (1 layer) |
summarize + frozen snapshots |
Docker |
self-improving skills |
Self-improving skills, FTS5 search, frozen memory snapshots |
coding-agent |
beta |
pip |
260,000 |
MIT |
Scan memory writes before persisting (anti-injection); Full session transcripts + FTS5 search alongside curated memory |
| oh-my-claudecode |
24.4k |
TypeScript |
plan-execute (file IPC, mkdir locking) |
-- |
-- |
-- |
none (Claude Code plugin) |
19-agent team via file IPC, Haiku-to-Opus model routing |
dev-tool |
beta |
npm |
194,000 |
MIT |
mkdir-based cross-process locking (atomic, stale detection); Model tier routing: Haiku for critic, Opus for deep reasoning |
| oh-my-codex |
19.6k |
TypeScript |
plan-execute (git worktree isolation) |
plugin process isolation (1 layer) |
-- |
git worktree |
MCP + hook plugin SDK |
30 agents, git worktree isolation, autoresearch loop |
dev-tool |
beta |
npm |
124,000 |
-- |
Plugin hook SDK with process isolation (1.5s timeout escalation); Heuristic task-to-worker allocation via file-path overlap; Autoresearch loops with score-based keep/discard |
| Guardrails AI |
6.6k |
Python |
validator pipeline (reask loop) |
none (it IS the security layer) |
-- |
-- |
Validator Hub (npm for constraints) |
Validator Hub, reask self-correction loop, 8 OnFailAction variants |
security-framework |
production |
pip |
18,079 |
Apache-2.0 |
OnFailAction pattern: 8 variants covering full error-handling tree; Streaming chunk accumulation at sentence boundary; Reask loop with budget ceiling |