| 1 |
SimonAgent |
Agent Runtime |
Python |
~200 |
"Zero frameworks, zero magic" — built from scratch in pure Python with a small, auditable core. 73 unit tests cover the loop, context, memory, and security branches. |
Dual-layer context compression (local sliding window + LLM compaction), on-demand skill loading, and a security boundary that blocks arbitrary command execution and restricts web access to public addresses. |
| 2 |
DragonScale |
Embedded Agent Runtime |
Rust/Go |
~500 |
A compact runtime for Linux and embedded environments that fixes the architectural gaps in single-binary agents: no privilege boundary, flat memory, and sequential tool calls. |
Capability-based security with an Isolated Tool Runtime, DAG-based parallel tool execution, MemGPT-style tiered memory (working, recall, archival, observational), and progressive tool disclosure via fuzzy search. |
| 3 |
tachi-agent |
Local-First Runtime |
TypeScript |
~300 |
A pluggable ReAct hub that fuses persistent memory (dokoro) with a multi-model council (TachiBot) over MCP — and defaults to 100% local inference via Qwen2.5 on Ollama. |
The agent loop lives in the client, never inside an MCP server, keeping the hub small and embeddable. Memory is a "bookend" by default (recall before, log after) but can be moved into the loop with a single flag. |
| 4 |
unimcp |
MCP Aggregator |
TypeScript |
~400 |
Solves MCP config duplication and tool visibility without server-side profiles. One config, one shared daemon, per-client tool control via environment variables. |
Client-side UNIMCP_INCLUDE / UNIMCP_EXCLUDE means each client declares its own tool surface. Auto-spawns on first use, shuts down after 30s idle, and hot-reloads on config change — no orphaned Docker containers or zombie stdio processes. |
| 5 |
DeepMCPAgent |
MCP Agent Framework |
Python |
~800 |
A plug-and-play framework for dynamic MCP tool discovery that works with any LLM and connects to multiple MCP servers simultaneously. |
Tools are typed through a three-stage pipeline (JSON-Schema → Pydantic → LangChain BaseTool), making MCP tools first-class citizens in LangGraph. Supports HTTP/SSE servers with headers and auth, and ships with a CLI for interactive agent chat. |
| 6 |
PicoClaw (Sipeed) |
Ultra-Light Runtime |
Go |
~2k |
An independent Go implementation of a personal AI assistant that runs in less than 10 MB of RAM and can be deployed on a Raspberry Pi. |
95% of the core code was agent-generated through a self-bootstrapping process, making it both a functional runtime and a case study in autonomous AI development. |
| 7 |
NanoBot |
Educational Runtime |
Python |
~1.5k |
Roughly 4,000 lines of pure Python implementing over 90% of OpenClaw's core capabilities — WebUI, tools, long-term memory, MCP integrations, model routing, and multi-agent delegation. |
The clearest codebase for learning how an agent harness actually works. pip install ready, no compilation, and a structure that maps directly to the conceptual components of an agent loop. |
| 8 |
ZeroClaw |
Zero-Overhead Runtime |
Rust |
~1.2k |
A single-binary Rust runtime with a memory footprint of ~7.8 MB. Four traits (Provider / Channel / Memory / Tool) define the entire architecture. |
Trait-driven design means every component is swappable. Connects to ~20 LLM providers, runs on Discord/Telegram/Slack, and has near-instant cold starts. The compilation requirement is the only barrier. |
| 9 |
magg |
MCP Aggregator |
Python |
~600 |
A meta-MCP server that aggregates multiple MCP services into a single unified endpoint — essentially giving AI assistants the power to extend their own capabilities on demand. |
Minimal dependencies, lightweight proxy architecture, and a clean separation between configured MCP sources and the unified interface presented to clients. A simpler alternative to unimcp for Python-native stacks. |
| 10 |
HarnessRouter |
Multi-Harness Router |
Python |
~700 |
The self-hosted, Apache-2.0 unified interface for running multiple agent harnesses behind one API. Instead of committing to one harness, you route tasks to whichever harness fits. |
Solves the harness lock-in problem directly. Run Claude Code, Codex, and Gemini CLI side-by-side with a single control plane. Still early, but the architectural premise is sound. |
| 11 |
emergent-wisdom/orchestrator |
Multi-Agent Engine |
Python |
~300 |
A lightweight multi-agent orchestration engine built around an "Inbox Zero" architecture and stigmergic collaboration — agents coordinate through a message bus rather than a central planner. |
Three orchestration modes (queue-based, two-phase, four-phase pipelines) with native MCP support. The stigmergic model — where agents leave traces in the environment that influence others — is a genuinely different coordination primitive. |
| 12 |
stagechain |
Pipeline Engine |
Python |
~200 |
A generic engine for chaining and routing tasks between AI agents — or any callable steps at all. It was built to demonstrate orchestration and pipeline-design patterns as reusable infrastructure, not as a wrapper around a specific application. |
Framework-agnostic by design. You can mix LLM calls, deterministic functions, and agent invocations in the same pipeline. Useful for teams that want orchestration without adopting a full agent framework. |
| 13 |
brAIn |
Event-Triggered Orchestration |
Python |
~150 |
A node-based agent framework where orchestration is event-triggered, not scheduler-driven. If your agent's cadence is "every Monday morning summarize X," brAIn treats the schedule as a primitive rather than bolting it on. |
Reactive many-to-many nodes with criticality-based preemption. The event-first model is a better fit for ambient agents that respond to signals rather than run on timers. |
| 14 |
niflheim-x |
Lightweight Orchestration |
Python |
~250 |
A composable agent orchestration framework positioned as "the fast alternative to LangChain." Installable via pip3 install niflheim-x. |
Focused on reducing the overhead of LangChain-style abstractions while preserving composability. A good fit for teams that found LangChain too heavy but still want structured orchestration. |
| 15 |
JL-Engine-Local |
Dynamic Assembly Engine |
— |
~200 |
A dynamic agent-assembly engine that builds and runs AI agents entirely in RAM, wiring up tools and behavior on the fly. No persistent agent definitions — the agent is assembled per task. |
The in-RAM assembly model makes it possible to create ephemeral agents with precise tool surfaces for each invocation, reducing context bloat and eliminating state leakage between tasks. |
| 16 |
Honcho Self-Hosted |
Memory Layer |
Python |
~400 |
A self-hosted deployment of Honcho, the cross-session memory layer that powers Hermes Agent's L4 memory. Runs the full stack (API, Deriver, PostgreSQL, Redis) on your own machine with no code changes to Hermes. |
Data sovereignty without sacrificing memory quality. Routes LLM calls through any OpenAI-compatible provider, and the tradeoff versus Plastic Labs' Neuromancer models is well-documented (slightly less precise deductive reasoning, same observation extraction). |
| 17 |
runtimenoteslabs/memory-layer |
Outcome-Based Memory |
Python |
~300 |
Persistent memory for coding agents with outcome-based learning. It doesn't just store what happened — it learns which memories actually help by tracking whether recalled advice led to successful outcomes. |
The outcome-based learning loop is the differentiator. Most memory layers store and retrieve; this one closes the feedback loop between memory retrieval and task success. |
| 18 |
nautilus-compass |
Black-Box Memory |
TypeScript |
~250 |
The only public memory layer that doesn't burn LLM tokens to extract facts before storing. It treats memory as a black box and detects drift without an extraction step. |
Eliminates the token cost of memory ingestion entirely. For agents that run frequently and generate large conversation logs, this is a meaningful cost reduction. |
| 19 |
MemoryLayer |
Verified Memory |
— |
~200 |
Local-first AI memory with reasoning and formal verification. It reasons over what it's told, catches contradictions, derives facts you never stated, and checks every claim before your agent repeats it. |
Drops into Claude Code, Cursor, and Windsurf via one MCP config line. The formal verification layer is unusual in the memory space — most memory layers trust what they store. |
| 20 |
oh-my-pi / can1357 |
Coding Agent |
TypeScript |
~2.6k |
A terminal coding agent with hash-anchored edits, an optimized tool harness, LSP integration, and subagent support. It doesn't try to be a platform — it tries to be a very good terminal agent. |
Hash-anchored edits solve a real problem in coding agents: diffs that break when line numbers shift. The LSP integration gives the agent semantic understanding of the codebase rather than just text. |
♠️ Critical Nexus ✊🏿
reddthat.com
PumpkinDrama@reddthat.com