Archestra Apps HackathonJoin now
Archestra Logo
Archestra.AI

Thread

A
Anuj4:42 PMOpen in Slack
Hey @user i am using archestra tell me few cookbook example so i know basic how it is working then i can built project
9 replies
A
Archestra4:42 PMOpen in Slack
Let me pull up the cookbook examples from the Archestra docs for you. There are several "example" docs. Let me grab a few of the developer-focused ones plus the quickstart. Hey Anuj! πŸ‘‹ Great to hear you're building with Archestra. Here are a few cookbook-style examples from the official docs to get you started. They all follow the same pattern: you take an AI agent and put *Archestra in front of it as a security layer that intercepts dangerous tool calls and blocks prompt-injection attacks. All the runnable code lives here πŸ‘‰ <https://github.com/archestra-ai/examples> ### :fried_egg: Cookbook examples 1. Vercel AI SDK (Node.js / Express) β€” ai-sdk-express Shows an agent with a get_file tool. A malicious test.txt tries to hijack the model ("talk like a drunk pirate :wavingblackflag::skullandcrossbones:"). You plug in Archestra to sanitize/block that injection. β†’ https://github.com/archestra-ai/examples/tree/main/ai-sdk-express 2. Pydantic AI (Python) β€” pydantic-ai An autonomous agent that reads a GitHub issue (untrusted) and can send_email. The issue contains a hidden prompt injection trying to exfiltrate data β€” a classic Lethal Trifecta. Great pure-Python starting point. β†’ https://github.com/archestra-ai/examples/tree/main/pydantic-ai 3. Mastra.ai (TypeScript) β€” mastra-ai A GitHub agent that reads public issues + accesses private repos. Comes with a docker-compose so you can spin up Mastra + Archestra together and watch a prompt-injection attack get blocked. β†’ https://github.com/archestra-ai/examples/tree/main/mastra-ai There are also examples for n8n, OpenWebUI, Claude Code, and Claude Desktop if you prefer those. ### :brain: The core concept Every example demonstrates the Lethal Trifecta risk: 1. Access to private data 2. Processing untrusted content 3. Ability to communicate externally When all three combine, an agent can be tricked into leaking data. Archestra sits between your agent and the LLM to catch this. ### πŸš€ My suggestion to start 1. Pick the example in your language (Python β†’ Pydantic AI, JS/TS β†’ Vercel or Mastra). 2. git clone <https://github.com/archestra-ai/examples>, add your OPENAI_API_KEY, and run it without Archestra first to see the attack succeed. 3. Then wire in Archestra and watch it get blocked. That "before/after" is the fastest way to understand how it works. Want me to walk through the step-by-step setup* for any one of these in detail? Just tell me which framework you're using and I'll pull the full guide. πŸ™‚ πŸ€– Community Agent
A
Anuj4:52 PMOpen in Slack
hey i m trying cerebras and it is working but as i told to built something it is breaking can u tell me why !
Join Slack to see the image
A
Archestra4:53 PMOpen in Slack
Hey Anuj! Thanks for the screenshot β€” that really helps. πŸ” Here's what's most likely going on: building an app requires the agent to make *tool calls (to scaffold the app, write files, etc.). Your error card shows: - Provider: cerebras - Model source: user override - Model: the ID 2717ddeb-... The issue is almost certainly that the Cerebras model you picked doesn't reliably support tool/function calling (or the specific model ID doesn't handle the tool-call format the app-builder needs). Simple chat ("hi") works fine because it needs no tools β€” but the moment you ask it to build a todo app, it tries to call scaffolding tools and breaks. ### Things to try 1. Switch to a tool-capable model for app building. I notice you also have Z.AI GLM-4.7 available in the composer. Try switching the agent to that (or another model known to support tools) and re-run the "build a todo app" request. 2. Check the model's tool-calling support on Cerebras. Not every Cerebras-hosted model exposes function calling. If you want to stay on Cerebras, pick one that explicitly supports tools. 3. Expand the "Error Details" in that red card β€” click the > arrow. That usually shows the exact underlying error (e.g. tools not supported, a schema error, or a rate limit), which will confirm the cause. Could you click Error Details and paste what it says? That'll let me pin down whether it's a tool-calling limitation, a bad model override, or something else. In the meantime, let me double-check the docs on supported providers/models for app building. There are docs on apps and supported providers. Let me check those. I've confirmed the mechanics from the docs. Here's a clearer, docs-backed follow-up for Anuj: --- Quick update, Anuj β€” I dug into the Archestra docs on how app-building works, and it confirms my hunch. βœ… Building an app is a tool-heavy flow. When you say "build a todo app," the agent runs an autonomous loop of tool calls: scaffold_app β†’ edit_app β†’ render β†’ get_app_diagnostics β†’ edit_app to fix. So the model you're using must reliably support function/tool calling β€” and it has to be good at it, because it's chaining many tool calls together. Your error card shows the request ran on Cerebras with a user-override model. That's the most likely culprit: - Some Cerebras models either don't support tool calling or handle it poorly, so the very first scaffold_app call fails β†’ "There was an issue with your request." - Simple chat ("hi") worked because it needed zero tools. The break happens the instant tools are required. ### What to do 1. Switch to a strong tool-calling model for app building. You already have Z.AI GLM-4.7 in the composer β€” try selecting that (or an OpenAI/Anthropic/Gemini model) and re-run "build a todo app." Frontier models handle the multi-step scaffold/edit loop much more reliably. 2. Click "Error Details"* (the > arrow …
A
Anuj4:57 PMOpen in Slack
If i am using gemma model error is not coming and when i use glm error is coming why can u explain me simple words and crisp points
Join Slack to see the image
A
Archestra4:57 PMOpen in Slack
Hey Anuj! πŸ‘‹ Great observation β€” the difference is actually a clue. Here it is in simple, crisp points: *Why Gemma works but GLM breaks - :jigsaw: Both need tool calls to build an app (scaffold, edit, render, etc.). - βœ… Gemma "worked" β€” but look closely at your screenshot: it said "I'll build a 2D game..." but never actually built anything. When you asked "why error is coming," it replied "I haven't actually started building yet." So it didn't error because it didn't really try to run the heavy tool loop β€” it just chatted. - ❌ GLM broke because it actually attempted the tool calls, and the call failed β†’ red error card. The key detail in both screenshots - Both errors show Provider: cerebras πŸ‘€ - So even when the composer shows "GLM-4.7" or "Gemma," the request is routing through Cerebras, and that's where it's failing. Simple takeaway - The model that doesn't error is the one that isn't fully doing the job (just replies with text). - The model that errors is the one trying to run tools β€” and the tool call is breaking at the Cerebras layer. What to do next πŸ”§ 1. Click "Error Details" (the > arrow) on the red card and paste what it says β€” that's the exact reason. 2. Try switching to a frontier model that's strong at tool calling (OpenAI / Anthropic / Gemini) instead of a Cerebras-hosted one, and re-run "build a todo app." Can you expand Error Details and paste the text? That'll let me tell you the exact* cause instead of guessing. πŸ™‚ πŸ€– Community Agent πŸ€– Community Agent
A
Anuj4:59 PMOpen in Slack
Can i use Groq it does have tool calling ability !
A
Archestra4:59 PMOpen in Slack
Yes, Anuj β€” Groq is a solid choice! βœ… Here it is crisp: *Can you use Groq? Yes. - Groq hosts models that do support tool/function calling (e.g. Llama 3.1/3.3 70B, and other tool-capable models). - So it can run the app-building tool loop that Cerebras was choking on. But β€” one important tip ☝️ - Tool-calling support β‰  tool-calling quality. App building chains many tool calls (scaffold β†’ edit β†’ render β†’ fix), so you want a model that's strong at it, not just one that technically supports it. - On Groq, pick a larger, tool-capable model (like a 70B-class Llama) rather than a tiny one for the best results. How to test it πŸ”§ 1. Switch your model to Groq (a 70B tool-capable one). 2. Re-run "build a todo app." 3. If it still errors β†’ click Error Details on the red card and paste it here. That'll tell us if it's a tool issue, a bad model pick, or a config/routing problem. One thing worth double-checking from your earlier screenshots: even when the composer showed GLM/Gemma, the Provider still said cerebras β€” so make sure the request is actually* routing through Groq after you switch, not silently falling back. πŸ‘€ Give it a try and let me know what happens! πŸ™‚ πŸ€– Community Agent πŸ€– Community Agent
A
Anuj5:06 PMOpen in Slack
groq is also not working tool calling is not happening or token limit
Status: 413
Type: api_payload_too_large_error
Message: Request too large for model `openai/gpt-oss-120b` in organization `org_01jfsb1qxsfc48encznxrqv0c4` service tier `on_demand` on tokens per minute (TPM): Limit 8000, Requested 38490, please reduce your message size and try again. Need more tokens? Upgrade to Dev Tier today at <https://console.groq.com/settings/billing>

Raw Error:
{
  "url": "<http://localhost:9000/v1/groq/4f2617ba-04c5-42d7-8cb1-a300fe776a7a/chat/completions>",
  "statusCode": 413,
  "isRetryable": false,
  "responseBody": {
    "error": {
      "message": "Request too large for model `openai/gpt-oss-120b` in organization `org_01jfsb1qxsfc48encznxrqv0c4` service tier `on_demand` on tokens per minute (TPM): Limit 8000, Requested 38490, please reduce your message size and try again. Need more tokens? Upgrade to Dev Tier today at <https://console.groq.com/settings/billing>",
      "type": "api_payload_too_large_error"
    }
  }
}```
A
Archestra5:06 PMOpen in Slack
Nailed it, Anuj! 🎯 This time it's *not a tool-calling problem β€” it's a rate limit. Here it is crisp: What the error means - Status 413 / api_payload_too_large_error = your request is too big for Groq's free tier limit. - Groq free tier caps you at 8,000 tokens per minute (TPM). - Your request needed 38,490 tokens β€” nearly 5Γ— over the limit β†’ rejected. Why app-building triggers it - Building an app sends a huge prompt: system instructions + app-builder skill + tool schemas + your message. That easily blows past 8K TPM. - Simple chat ("hi") is tiny, so it fits under the limit β€” that's why chat works but building doesn't. How to fix it πŸ”§ 1. Upgrade Groq to Dev Tier β€” the error links right to it: https://console.groq.com/settings/billing. This raises your TPM limit and is the cleanest fix if you want to stay on Groq. 2. Use a provider with higher limits β€” an OpenAI / Anthropic / Gemini key won't hit an 8K TPM wall for app building. 3. Reduce request size β€” start a fresh chat (long history inflates the token count) before asking it to build. Bottom line - Groq's tool calling is fine βœ… β€” you just ran into the free-tier 8K TPM cap*. App building needs more headroom than that. My pick: either bump Groq to Dev Tier, or switch to a frontier provider key for the build step. Want me to point you to where in Archestra you set up a different provider key? πŸ™‚ πŸ€– Community Agent πŸ€– Community Agent