Awesome agentic engineering

A curated list of tools, frameworks, and practices for building AI agents and agentic pipelines: agent orchestration, the Model Context Protocol (MCP), prompt compilation, critic/reviewer patterns, spec-driven agent development, local-first LLM tooling, and cost-tracked AI-agent workflows.

This list follows the sindresorhus/awesome convention (badge, Contents, categorized entries, Contributing, License), verified live against that repo's own README and pull_request_template.md on 2026-09-07. See docs/MIPs/MIP-0043-awesome- agentic-engineering-list.md §4.1 for what was checked. It is maintained from marola, the ocean-intelligence Telegram assistant this repo belongs to. See the Top 10 repos most similar to marola section below for how marola itself fits this landscape.

Contents

• Agent Frameworks & Orchestration
• MCP Tooling
• DSPy & Prompt Compilation
• Critic / Reviewer-Pattern & Multi-Agent Pipelines
• Spec / RFC-Driven Agent Dev-Loops
• Local-First / Ollama-Based Agents
• Cost- and Usage-Tracked Agent Development
• Top 10 GitHub repos most similar to marola

Agent Frameworks & Orchestration

• LangChain - A widely used framework for building LLM-powered applications, chains, and agents across many model providers.
• LangGraph - A graph-based runtime for building stateful, multi-step, multi-agent LLM applications with explicit control flow.
• AutoGen - Microsoft's framework for building multi-agent conversational AI applications, including agent-to-agent collaboration patterns.
• CrewAI - A framework for orchestrating role-playing, autonomous AI agents that collaborate on tasks as a "crew."
• OpenHands - An open-source platform for AI agents that can modify code, run commands, browse the web, and call APIs to perform software-engineering tasks.
• llm4s - A Scala framework for LLM applications emphasizing type safety and functional programming, with multi-provider clients, agent/tool-calling, RAG, and observability on the JVM: one of a small number of Scala/JVM-native (not Java-first) agent frameworks.

MCP Tooling

• Model Context Protocol servers - The official reference-implementation repo for MCP servers (Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time), the canonical pattern any app exposing its own capabilities as an MCP tool server follows.
• mcp-client-for-ollama - A terminal UI connecting local Ollama models to MCP servers, with agent mode, multi-server support, and human-in-the-loop controls, entirely local.
• open-meteo-mcp - An MCP server exposing Open-Meteo's weather APIs, including its Marine Weather endpoint (wave height/period/direction, sea surface temperature), to LLMs.

DSPy & Prompt Compilation

• DSPy - Stanford NLP's framework for programming (not manually prompting) language models, with optimizers that compile modules into tuned prompts and few-shot demonstrations, decoupling program logic from prompt text.

Critic / Reviewer-Pattern & Multi-Agent Pipelines

• Self-Refine - The reference implementation of the "Self-Refine" paper: an LLM generates output, produces feedback/critique on its own output, then revises: an explicit generate → feedback → refine loop, evaluated across acronym generation, dialogue, code readability, and math.

Spec / RFC-Driven Agent Dev-Loops

• Spec Kit - GitHub's open-source toolkit for spec-driven development with coding agents: specify → plan → tasks → implement. Writing what to build before building it, as a structured workflow rather than a single prompt.

Local-First / Ollama-Based Agents

• Ollama - Run large language models locally, with a simple CLI and API: the local-model runtime underlying most "no cloud account required" agent stacks.
• mcp-client-for-ollama - See MCP Tooling above: Ollama models driving MCP tool calls, fully local.

Cost- and Usage-Tracked Agent Development

• ccusage - A CLI that parses local Claude Code/Codex/ OpenCode session logs into daily/weekly/session cost and token reports, making AI-agent development spend visible and attributable per unit of work rather than left implicit.

Top 10 GitHub repos most similar to marola

Hand-researched and verified live on 2026-09-07 (each repo's real GitHub page fetched; star counts approximate as of that date; see docs/MIPs/MIP-0043-awesome-agentic-engineering-list.md for the MIP behind this doc). "Similar" is scored on three axes: domain (environmental/ocean/weather data agents), architecture (local-first Ollama multi-agent pipelines, critic/reviewer-pattern agents, DSPy-based prompt compilation, MCP-server-exposed apps), and philosophy (design-doc- before-code culture, AI-agent-run/AI-native repos, cost-tracked agent development). Each entry states explicitly which axis(es) it matches and why.

1. stanfordnlp/dspy (~37.8k★): Stanford NLP's framework for programming (not manually prompting) LMs, with optimizers that compile modules into tuned prompts/few-shot demos. Architecture axis: this is the exact upstream project marola's offline dspy/ compile step is built on: the same "compile once, replay the artifact at runtime" idea marola applies by loading a JSON artifact from Scala instead of keeping a runtime DSPy dependency.
2. modelcontextprotocol/servers (~90.1k★): the official reference-implementation repo for MCP servers. Architecture axis: marola's cli/ module exposes its own capabilities as an MCP tool server (just mcp-server); this repo is the canonical pattern/spec marola's server follows.
3. getkyo/kyo (~812★): a Scala 3 toolkit built on algebraic effects (the A < S pending type), where an unhandled error or undeclared effect fails to compile, targeting JVM/JS/Native/Wasm. Architecture axis: this is literally the pre-1.0 effects library marola is built on (core/cli use Kyo directly), not just similar: it's the same dependency.
4. cmer81/open-meteo-mcp (~68★): an MCP server exposing Open-Meteo's weather APIs, including its Marine Weather endpoint (wave height/period/ direction, sea surface temperature). Domain + architecture axes: same data source marola consumes (Open-Meteo) for the same marine/sea-conditions purpose, wrapped in the same MCP-server pattern marola itself uses to expose its own tools.
5. github/spec-kit (~133.8k★): GitHub's own toolkit for spec-driven development with coding agents: specify → plan → tasks → implement. Philosophy axis: near-exact structural match to marola's MIP-driven, design-doc-before-code culture (docs/MIPs/, the mip/mip-tasks skills, docs/3-Working-on-the-repo/DEV-FLOW.md's idea→MIP→tasks→PR loop); both make a written spec/plan a mandatory gate before implementation.
6. jonigl/mcp-client-for-ollama (~816★): a terminal UI connecting local Ollama models to MCP servers, with agent mode, multi-server support, and human-in-the-loop controls, entirely local. Architecture axis: directly parallels marola's local-first design: Ollama as the model backend plus MCP as the tool- exposure layer, with zero required cloud account.
7. llm4s/llm4s (~257★): a Scala framework for LLM applications emphasizing type safety and FP, with multi-provider clients, agent/tool-calling, RAG, and observability on the JVM. Architecture axis: one of the very few Scala/JVM-native (not JVM-via-Java-first) LLM agent frameworks, the same ecosystem niche marola's Scala 3 + Kyo pipeline occupies (see docs/4-Research-and-plans/AGENT-FRAMEWORKS-SURVEY.md §1.2 and MIP-0012).
8. ryoppippi/ccusage (~18.4k★): a CLI that parses local Claude Code/Codex/OpenCode session logs into daily/weekly/session cost and token reports. Philosophy axis: matches marola's Cost: git-trailer discipline and just cost-split/ cost-fill tooling; both are explicit, tool-enforced attempts to make AI-agent development spend visible and attributable per unit of work rather than left implicit.
9. madaan/self-refine (~820★): the reference implementation of the "Self-Refine" paper: an LLM generates output, produces feedback/critique on its own output, then revises. Architecture axis: the closest published academic match to marola's summarizer-then-Reviewer two-pass pattern, where a second LLM pass explicitly grades/ corrects the first.
10. open-meteo/open-meteo (~6.2k★): the open-source weather API/service itself (AGPLv3, free for non-commercial use), aggregating NOAA/DWD/ECMWF/JMA models, with a dedicated Marine Forecast API. Domain axis: the actual upstream data provider marola's sea/weather-conditions pipeline is built on, and its "free for non-commercial use, self-hostable, zero mandatory account" ethos mirrors marola's own "zero-cloud-account-required, everything opt-in" design stance.

Ruled out during research (kept here for honesty, not padding): calimero-network/ai-code-reviewer (real, ~9★, but a parallel multi-model voting reviewer, not a sequential generate→critique pattern; self-refine is the closer match); Jellyfish-AI/jellyfish-mcp (real, but an MCP server for the Jellyfish.ai engineering-analytics SaaS: a naming coincidence, not a marine-jellyfish domain match); no public GitHub repo was found combining "marine-jellyfish detection + Telegram bot + LLM agent" as one verifiable project, so none was forced into the list.

How this list is kept updated

scripts/awesome_agentic_digest.py queries the real GitHub Search API (api.github.com/search/repositories, no auth needed for reasonable unauthenticated use) across verified agentic-engineering topics (topic:agents, topic:llm-agents, topic:ai-agents, topic:mcp, topic:multi-agent-systems), sorted by stars and by recent activity, and caches candidates under .tmp/awesome_agentic_cache/ (gitignored). It never writes to this file. Run it, review the printed/cached candidates by hand, open the ones that look real and genuinely relevant, and hand-write an entry in the matching section above, in this doc's own - [Title](URL) - Description. format, the same human-gated "propose, don't auto-merge" pattern docs/MIPs/MIP-0041-soft-book-ingestion.md established for book-derived rule changes. See docs/MIPs/MIP-0043-awesome-agentic-engineering-list.md for the full design and verification.

python3 scripts/awesome_agentic_digest.py            # fetch, cache, print a human-readable summary
python3 scripts/awesome_agentic_digest.py --json      # machine-readable summary
python3 scripts/awesome_agentic_digest.py --self-test # no network, run via `just quality-other`

Contributing

This list is maintained inside the marola repository. To propose an addition: run scripts/awesome_agentic_digest.py (above) to find candidates, or suggest one you already know of directly: either way, a human confirms the repo genuinely exists and fits a category before it's added, following the sindresorhus/awesome entry format: - [Title](URL) - Description., description starting uppercase and ending with a period, describing the project itself rather than the list, no marketing language.

License

This list's own curated text (titles, descriptions, categorization) is released under CC0 1.0, following the sindresorhus/awesome convention. The license applies only to this document's curation text, not to the licenses of the linked projects themselves.