Active
Agent toolkit
Pi coding agent
Earendil Inc. + contributors
A small, hackable coding loop with reusable agent, TUI, and provider libraries.
- Surface
- CLI + SDK
- Implementation
- TypeScript
- Model connection
- 15+ providers + custom models
- License
- MIT
OPEN AGENT INFRASTRUCTURE · SOURCE-LED DIRECTORY
The model is only half the system.
Coding agents, execution harnesses, and orchestration frameworks determine how a model sees a repository, calls tools, applies patches, and recovers from failure. This catalog keeps the scaffold visible.
17
open projects
10
ready-to-run agents
6
build frameworks
1
compatibility map next
READY-TO-RUN AGENTS
Put an agent inside the repository.
These projects provide an opinionated working loop: collect context, call a model, use tools, change files, run commands, and return evidence to a human.
Active
Coding agent
Codex CLI
OpenAI
Local repository work with approvals, sandboxing, tool use, and MCP extensions.
- Surface
- CLI + IDE
- Implementation
- Rust
- Model connection
- OpenAI API + configurable providers
- License
- Apache 2.0
Active
Coding agent
OpenCode
Anomaly
A provider-flexible coding agent with terminal, desktop, and editor surfaces.
- Surface
- CLI + desktop + IDE
- Implementation
- TypeScript
- Model connection
- Multi-provider + OpenAI-compatible APIs
- License
- MIT
Active
Pair programmer
Aider
Aider AI
Git-native pair programming, repository maps, controlled edits, linting, and tests.
- Surface
- CLI + editor watch mode
- Implementation
- Python
- Model connection
- LiteLLM + local/OpenAI-compatible models
- License
- Apache 2.0
Active
Coding agent
Qwen Code
Qwen / Alibaba
Terminal-first coding optimized for Qwen while retaining bring-your-own-provider paths.
- Surface
- CLI + IDE integrations
- Implementation
- TypeScript
- Model connection
- Qwen + OpenAI/Anthropic/Gemini-compatible APIs
- License
- Apache 2.0
Transitioning
Coding agent
Gemini CLI
Gemini-native terminal workflows, MCP, extensions, and enterprise Google Cloud access.
- Surface
- CLI
- Implementation
- TypeScript
- Model connection
- Gemini API + Google Cloud
- License
- Apache 2.0
Active
General agent
Goose
Agentic AI Foundation
Local code and knowledge-work automation with broad provider and extension support.
- Surface
- Desktop + CLI + API
- Implementation
- Rust
- Model connection
- 15+ providers + ACP + MCP
- License
- Apache 2.0
Active
Coding agent
Cline
Cline
Human-approved edits and commands across IDE, terminal, headless, and SDK workflows.
- Surface
- IDE + CLI + SDK
- Implementation
- TypeScript
- Model connection
- Multi-provider + local models
- License
- Apache 2.0
Active
Code checks + agent
Continue
Continue Dev
Source-controlled AI checks and repeatable repository policy in continuous integration.
- Surface
- CLI + CI + IDE
- Implementation
- TypeScript
- Model connection
- Multi-provider + OpenAI-compatible APIs
- License
- Apache 2.0
Active
Software agent platform
OpenHands
OpenHands
Composable software agents that can scale from one-off repo tasks to many sandboxes.
- Surface
- CLI + SDK + web
- Implementation
- Python + TypeScript
- Model connection
- Multi-provider SDK + local or sandboxed workspaces
- License
- MIT core
RESEARCH HARNESS
Keep the scaffold small enough to inspect.
A minimal harness is useful when the model—not a large product surface—should remain the variable under study.
Active
Research harness
mini-SWE-agent
Princeton + Stanford
A minimal, inspectable baseline for issue solving, evaluations, fine-tuning, and RL.
- Surface
- CLI + Python
- Implementation
- Python
- Model connection
- LiteLLM + OpenRouter + compatible endpoints
- License
- MIT
BUILD-YOUR-OWN FRAMEWORKS
Design the loop instead of adopting one.
These libraries expose the components behind an agent: model clients, tools, state, handoffs, graphs, approvals, sandboxes, tracing, and evaluation.
Active
Agent framework
PydanticAI
Pydantic
Typed production agents, structured outputs, tool approval, durable execution, and evals.
- Surface
- Python SDK
- Implementation
- Python
- Model connection
- Broad providers + Ollama + LiteLLM + Hugging Face
- License
- MIT
Active
Orchestration framework
LangGraph
LangChain
Long-running, stateful agent graphs with persistence and human intervention.
- Surface
- Python + JavaScript SDKs
- Implementation
- Python + TypeScript
- Model connection
- LangChain model integrations
- License
- MIT
Active
Agent framework
smolagents
Hugging Face
Compact code agents and tool-calling agents with a deliberately small abstraction layer.
- Surface
- Python SDK + CLI
- Implementation
- Python
- Model connection
- Hugging Face + API providers + local models
- License
- Apache 2.0
Active
Multi-agent framework
OpenAI Agents SDK
OpenAI
Tools, handoffs, guardrails, tracing, sessions, and sandboxed long-horizon agents.
- Surface
- Python + TypeScript SDKs
- Implementation
- Python + TypeScript
- Model connection
- OpenAI APIs + 100+ models through adapters
- License
- MIT
Active
Multi-agent framework
Microsoft Agent Framework
Microsoft
Production-grade agent orchestration across Python and .NET with MCP and A2A.
- Surface
- Python + .NET SDKs
- Implementation
- Python + C#
- Model connection
- Microsoft Foundry + OpenAI + extensible providers
- License
- MIT
Active
Agent application framework
Mastra
Mastra
TypeScript agents, graph workflows, model routing, memory, and human-in-the-loop apps.
- Surface
- TypeScript SDK
- Implementation
- TypeScript
- Model connection
- 40+ model providers
- License
- Apache 2.0 core
Inclusion rule: public source, a stated open-source license, and a runnable project or library. Closed clients and source-visible but non-open tools are excluded. Lifecycle labels are point-in-time signals, not endorsements.
MODEL × HARNESS · NEXT
Connect the two ledgers without pretending compatibility is binary.
The next layer should join each harness to each open model only when the interface, tool behavior, context controls, and operating requirements are explicit.
API fit
Native provider, OpenAI-compatible endpoint, LiteLLM, Ollama, vLLM, or a custom adapter.
Tool fit
Function-calling format, shell semantics, structured output, MCP, and agent-specific prompt contracts.
Context fit
Working context, compaction strategy, reasoning controls, cache behavior, and repository retrieval.
Operational fit
Hardware location, latency, sandboxing, approval policy, observability, and reproducible evaluation.