| name | aixplain-builder |
| description | Build, run, and deploy anything on the aiXplain platform with the Python SDK and REST/JS/OpenAI-compatible APIs — AI agents and multi-agent teams, direct model inference (LLMs, speech-to-text/Whisper, translation, vision, embeddings), knowledge bases and RAG, tools and integrations (Slack, Gmail, databases, MCP, custom Python functions), runtime governance/inspectors, memory, and serverless deployment. Use this whenever the user mentions aiXplain (aixplain/aiXplain), the `aixplain` SDK, `from aixplain import Aixplain`, `aix.Agent`/`aix.Model`/`aix.Tool`, an aiXplain API key, Studio, the marketplace, or asks to build/run/deploy an agent, transcribe audio, run a model, set up RAG, or wire an integration on aiXplain — even when they don't name every component explicitly. |
| metadata | {"requires":{"env":["AIXPLAIN_API_KEY"],"bins":["python3","pip"]}} |
aiXplain Builder
Design, run, and deploy everything the aiXplain SDK supports: agents and team agents, direct model inference, knowledge bases / RAG, tools and integrations, runtime governance, memory, and the access APIs. This skill is the single source of truth — every snippet here is verified against aiXplain SDK v0.2.44 (the unified v2 Aixplain client) and live SDK introspection.
Audience & self-containment
Used on behalf of both non-developers and developers. So:
- Do the technical work yourself. Write and run the code; don't hand the user a traceback. Diagnose it, fix it, and report in plain language: what happened, what you did, what's next.
- This skill is authoritative. Use its snippets as written; don't redirect the user to GitHub or the SDK source. If a snippet here fails against the installed SDK, the skill is wrong, not the user — fix it, then consider the "Report aiXplain issues" habit below.
- Lead with the result and the Studio link, not the implementation. Keep full code available for developers who want it.
Setup
Install/upgrade first, then initialize the client:
python3 -m pip install --upgrade aixplain
import os
from aixplain import Aixplain
aix = Aixplain(api_key=os.environ["AIXPLAIN_API_KEY"])
If AIXPLAIN_API_KEY isn't set, ask the user for it (or to add it to a .env) — get one at https://console.aixplain.com/settings/keys. Most operations spend credits (1 credit = $1); say so before running anything billable for a non-developer.
Pick the right tool for the job
| The user wants to… | Build a… | Reference |
|---|
| Reason over a goal, call tools, multi-step work | Agent | references/agents.md |
| Coordinate several specialists / a workflow | Team agent | references/agents.md |
| Just run a model once (LLM, transcription, translation, vision) | Direct model call | references/models.md |
| Answer from private documents | Knowledge base + agent (RAG) | references/knowledge-memory.md |
| Connect Slack/Gmail/DB/MCP, or wrap a Python function | Tool / integration | references/tools-integrations.md |
| Enforce a policy at runtime (safety, PII, format) | Inspector | references/governance.md |
| Remember across turns / sessions / agents | Session or shared memory | references/knowledge-memory.md |
| Call a deployed asset over HTTP/JS/OpenAI client | REST / access API | references/deployment-access.md |
| Update, or export-as-code, an existing/deployed agent | agent lifecycle | references/agents.md |
| A ready-made recipe to adapt | — | references/patterns.md |
Read the relevant reference file before writing code for that domain — they hold the exact signatures, IDs, and gotchas.
Not in the v2 SDK: pipelines, fine-tuning, benchmarking, and datasets/corpora. These are legacy-v1 only or Studio-only. See references/deployment-access.md § What the v2 SDK does NOT cover. For multi-step workflows in v2, use a team agent.
Workflow for building an agent
- Plan, then confirm. State the agent's name, description, instructions, which tools/integrations, and single vs. team. Wait for approval before building anything billable.
- Search before hardcoding.
aix.Tool.search(...), aix.Model.search(...), aix.Integration.search()["results"]. Never say "not available" without searching. IDs in the reference tables can drift — if one 404s, search by name.
- Create tools and scope them. For every tool, narrow
allowed_actions to the minimum the task needs — the default (all actions) over-privileges the agent and hurts its reasoning. See references/tools-integrations.md.
- Build the agent. Omit
llm unless the user asks for a specific model (the platform default is good). Set output_format and, for JSON, expected_output.
- Save = deploy.
agent.save() (or team.save(save_subcomponents=True)) promotes DRAFT → ONBOARDED and gives a persistent endpoint. There is no separate deploy().
- Run and verify.
agent.run(query=...), read .data.output. Inspect .data.steps if it misbehaves.
- Share Studio links so the user can edit/monitor visually:
- Builder/traces:
https://studio.aixplain.com/build/<AGENT_ID>/schema
- Analytics:
https://studio.aixplain.com/dashboard/analytics/?agent=<AGENT_ID>
Conventions that apply everywhere
- Resilient runs: sync via
.run(...); async via .run_async(...) then .sync_poll(url) (blocks, returns the same shape) or manual .poll(url) until .completed.
- Reading results: agents →
response.data.output, .status, .data.steps, .data.execution_stats; models → response.data, .status, .usage, ._raw_data.
- Unique names: if
save() raises name_already_exists, ask the user: update the existing asset or create under a new name. Appending int(time.time()) to a name keeps demos unique.
- Don't invent credentials or IDs. Ask the user for required inputs (DB URLs, API tokens). Use placeholders only when clearly labeled.
Passing local files to aiXplain
Models/agents/tools take URLs, not local paths. Upload first and pass the returned URL:
from aixplain.v2.file import FileUploader
url = FileUploader(api_key=os.environ["AIXPLAIN_API_KEY"]).upload(
"/path/to/file.mp3", is_temp=True, return_download_link=True)
Use return_download_link=True so the URL is browser-accessible. Size limits: audio 50 MB, image/documents 25 MB, video/database 300 MB. (For .html/.zip outputs the SDK's MIME detector can mislabel the extension — set the right content type when it matters.)
Report aiXplain issues (habit)
When you hit a quirk that is clearly aiXplain-caused (not the user's code/config) — an advertised action that always fails, a documented parameter the backend ignores, a connector requesting too few scopes — offer to file an issue at https://github.com/aixplain/aiXplain. Reproduce it first to confirm the cause is on aiXplain's side, draft the title/body, and get the user's approval before posting (the repo is public). Redact API keys, emails, and asset/account IDs; keep the repro generic. Include SDK version, asset type, a minimal repro, and expected vs. actual behavior.
External links
