惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

B
Blog RSS Feed
Jina AI
Jina AI
雷峰网
雷峰网
Blog — PlanetScale
Blog — PlanetScale
Hugging Face - Blog
Hugging Face - Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
Last Week in AI
Last Week in AI
博客园 - 司徒正美
罗磊的独立博客
J
Java Code Geeks
Engineering at Meta
Engineering at Meta
WordPress大学
WordPress大学
Vercel News
Vercel News
A
About on SuperTechFans
I
InfoQ
D
DataBreaches.Net
爱范儿
爱范儿
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
U
Unit 42
aimingoo的专栏
aimingoo的专栏
宝玉的分享
宝玉的分享
P
Proofpoint News Feed
Microsoft Azure Blog
Microsoft Azure Blog
美团技术团队

Hacker News - Newest: "LLM"

GitHub - lechmazur/position_bias: A benchmark for testing whether LLM judges keep the same preference when two lightly edited versions of the same story are shown in opposite orders. Flex routing (EU and EFTA) Dark Factories: Retooling for LLM Velocity Ask HN: What would be the impact of a LLM output injection attack? GitHub - Oaklight/llm-rosetta: Production-ready LLM API translation layer for Python — bidirectional conversion between OpenAI, Anthropic & Google formats via hub-and-spoke IR. Optional API gateway. Streaming & non-streaming. Zero core deps. Contributions welcome! GitHub - browser-use/browser-harness: Self-healing browser harness that enables LLMs to complete any task. GitHub - moeen-mahmud/remen: Remen turns thoughts into something you can return to Analyzing 156 LLM Launch Posts on Hacker News ChatGPT vs Gemini vs Claude: The Best LLM Subscription You Should Buy GitHub - salaamalykum/quran-semantic-search: High-density RAG Semantic Search Engine & Quran Corpus (GEO/SEO Architecture) GitHub - NVIDIA/TensorRT-LLM: TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT LLM also contains components to create Python and C++ runtimes that orchestrate the inference execution in a performant way. The State of LLM Bug Bounties in 2026 Operational Readiness Criteria for Tool-Using LLM Agents Meshcore: Architecture for a Decentralized P2P LLM Inference Network How an LLM becomes more coherent as we train it GitHub - seetrex-ai/laimark GitHub - Jossifresben/BibCrit: AI-assited biblical textual criticism GitHub - wastedcode/memex: File system based wiki, maintained by Claude 99helpers.com GitHub - cliver-project/AITrigram GitHub - unbody-io/adapt: A self-evolving memory layer for AI agents. GitHub - hb20007/awesome-gen-ai-fails: A list of incidents where reliance on generative AI and LLMs resulted in harm to companies, individuals, or society GitHub - nevenkordic/localmind: Run any local LLM with persistent memory and context. CLI agent over Ollama with SQLite-backed hybrid recall. No cloud. Ask HN: What are the machine requirements for a LLM like Llama-3.1-8B? Faster LLM Inference via Sequential Monte Carlo grpo explained: group relative policy optimization for llm finetuning - cgft Stop comparing price per million tokens: the hidden LLM API costs · TensorZero Andrej Karpathy's LLM Wiki Is a Bad Idea GitHub - GG-QandV/mnemostroma: Offline RAM-first cognitive leer/coprocessor for AI agents and robotics. Solves "Context Abandonment" with 20-80ms latency using a dual-thread biomimetic memory architecture (ONNX + SQLite WAL). mempalace/agent at agent · skorotkiewicz/mempalace
GitHub - msradam/ocarina: Automation framework for MCP se...
msradam · 2026-06-27 · via Hacker News - Newest: "LLM"
Ocarina

An automation framework for MCP servers. Write a YAML script, replay it deterministically, no LLM in the loop.

An Ocarina rondo building a 3D scene in Blender, one step at a time

A rondo driving blender-mcp: lay down a plane, drop a cube, stack a sphere, add a cone, then verify the scene. Same YAML, same result, every run. No model involved. Clone blender-mcp-ocarina to run it yourself.

The MCP ecosystem is already enormous: thousands of servers exposing real services through typed tools, readable resources, and schema-checked contracts, deployed and ready to call. Ocarina is an automation framework for all of it. Write a YAML script that drives tools across one or more servers, pipes values between steps, branches, loops, and retries, and runs the same way every time. No LLM in the loop, so every run is reproducible and costs nothing.

These tools were built to be read by language models, and language models are trained on human language, so the tools read cleanly to people too. A server exposes named contracts like get_issues or query_database, not endpoints you wire up in code. Every server someone built for an AI assistant is one you can drive.

What you write is a playbook: a portable artifact that captures an automation workflow over those servers, with MCP as the wire protocol. You can read it, review it in a pull request, version it, and run it anywhere the servers are reachable. Write it by hand or have an agent generate it. Either way it runs the same on every execution, with no sampling, no tokens, and nothing inferring between the file and the result.

Install

go install github.com/msradam/ocarina@latest

Binaries are available on the releases page. Building from source requires Go 1.26+.

Use

Generate markdown docs for a server:

ocarina docs uvx mcp-server-sqlite --db-path mydb.sqlite
ocarina docs npx -y @modelcontextprotocol/server-github > docs/github.md

Run a rondo:

ocarina play db-audit.yaml
ocarina play db-audit.yaml --dry-run
ocarina play db-audit.yaml -e db=/tmp/other.sqlite  # override a key at runtime

Validate a rondo against the live server without running any tools:

ocarina validate db-audit.yaml

Design principles

  • Deterministic. The same rondo produces the same result on every run. No sampling, no randomness.
  • Protocol-native. Talks MCP directly via tools/call, resources/read, and resources/list. Works with any compliant server.
  • Assertions are first-class. play exits non-zero if any expect: check fails. Rondos work as CI health checks out of the box.
  • No credentials in scripts. Server connection and environment variables stay outside the rondo file.
  • One rondo, any machine. If the MCP server is available, the rondo runs.

Rondo format

A rondo is a YAML file with three sections.

keys:
  owner: acme
  repo: api

server:
  command: npx
  args: [-y, "@modelcontextprotocol/server-github"]

rondo:
  - name: recent commits
    tool: list_commits
    args:
      owner: "{{owner}}"
      repo: "{{repo}}"
    grab: ".0.sha"
    echo: latest_sha

  - name: commit detail
    tool: get_commit
    args:
      owner: "{{owner}}"
      repo: "{{repo}}"
      sha: "{{latest_sha}}"
    expect:
      contains: "feat"

Step fields

Field Description
server Which server to run this step against (a key in the servers: map); defaults to the only/first server
tool Tool name to call
resource Resource URI to read (resources/read)
list_resources Server prefix to list resources from; output is a JSON URI array
args Tool arguments. {{key}} interpolates from keys or prior echo captures
echo Store this step's output under a key for later steps
grab Dot-path into JSON output before storing: .0.sha, .name, .items.0.id
loop Expand a JSON array key into repeated iterations; sets {{item}} each time
expect.contains Assert output contains this string
expect.matches Assert output matches this regex
expect.equals Assert output equals this string (whitespace-trimmed)
expect.is_error Assert whether the tool returned isError: true
ignore_errors Continue past failures instead of halting
tags Tag this step for --tags / --skip-tags filtering

{{env.NAME}} resolves from the process environment and works anywhere {{key}} does.

Coming from Ansible? tasks: is accepted as an alias for rondo:, and register: as an alias for echo:.

Multiple servers

A single rondo can talk to more than one server. Declare them under servers: and set server: on each step. Steps that omit server: use the first entry.

servers:
  time: {command: uvx, args: [mcp-server-time]}
  fetch: {command: uvx, args: [-y, "@modelcontextprotocol/server-fetch"]}

rondo:
  - name: get time
    server: time
    tool: get_current_time
    args: {timezone: UTC}

  - name: fetch page
    server: fetch
    tool: fetch
    args: {url: "https://example.com"}

Output and diff namespace tool names by server (time.get_current_time). The single server: block still works for one-server rondos.

Commands

ocarina docs <command> [args...]: generate markdown documentation for every tool, resource, and resource template a server exposes.

ocarina play <rondo.yaml>: execute each step against the live server.

ocarina validate <rondo.yaml>: check tool names, required args, schema types, and {{key}} data flow without making any calls.

ocarina hum <command> [args...] -- <tool> [key=value ...]: call a single tool and print the result.

ocarina record <output.yaml> <command> [args...]: proxy mode; records every tool call from a live MCP client session into a rondo file.

Server names

Create a .mcp.json (or ~/.mcp.json for credentials) and reference servers by name in rondos and on the command line:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    }
  }
}
ocarina hum github -- list_commits owner=pytorch repo=pytorch per_page=1

See mcp.json.example for a starter template. Ocarina also discovers servers from the Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json).

Examples

Working rondos for 50+ MCP servers are in examples/. A selection:

Rondo Server What it does
sqlite/data-quality-audit.yaml mcp-server-sqlite Schema check, row counts, referential integrity assertions
github-investigation/repo-health.yaml github-mcp-server Commit history, open issues, contributor activity
github-investigation/resource-audit.yaml github-mcp-server Read repo files directly via resource: steps
postgres/query-workflow.yaml mcp-server-postgres Multi-step query and result validation
docker/docker.yaml mcp-server-docker Container list, image audit, resource usage check
elasticsearch/cluster-search.yaml mcp-server-elasticsearch Index health, search, document count assertions
playwright-browser/page-audit.yaml mcp-server-playwright Navigate, extract content, assert on page state
yahoo-finance/portfolio-health.yaml mcp-yahoo-finance Price fetch, income statements, parameterized by ticker

See docs/tested-servers.md for the full list.

Showcases

Standalone repositories you can clone and run, each a real working environment for a different MCP server:

  • duckdb-mcp-ocarina: data integrity, migration, and regression tests against a DuckDB database. Clone and run, no credentials.
  • chrome-devtools-mcp-ocarina: synthetic web health checks through Google's Chrome DevTools MCP. Fail on a console error or a failed request.
  • github-mcp-ocarina: repo governance as tests through the GitHub MCP server. Assert a repo ships a license, is documented, and has history.
  • blender-mcp-ocarina: automate and snapshot-test a 3D scene in Blender, an app with no external API at all.

Use in CI

play exits 0 if all expect: assertions pass, non-zero otherwise. Drop a rondo into any CI pipeline:

- name: Database health check
  run: ocarina play rondos/db-audit.yaml

A composite GitHub Action installs Ocarina and replays a rondo:

- uses: msradam/ocarina@v1
  with:
    rondo: tests/mcp-smoke.yaml

See action.yml and .github/workflows/example.yml.

License

MIT. Whistle icon by Alessio Capponi from Noun Project (CC BY 3.0).