AI skills employers want: the tools AI job postings actually ask for

The AI skills employers want are mostly not AI tools. On Tokn, as of 12 October 2026, Python appears in half of the roughly 770 live AI job postings we track in Europe, followed by cloud platforms, Kubernetes and PyTorch. LLM frameworks such as LangChain show up in about 1 posting in 18. A less obvious requirement also shows up: proof that you work every day with AI coding agents such as Claude Code, Cursor or Codex.

Most "AI tools to learn" lists are written from vendor pages. We did it the other way round: we took every active posting on Tokn, collected from company career pages, and counted which tools they name and which roles ask for each.

In short

• Python (389 jobs at 61 companies) is the one near-universal requirement. Cloud and infrastructure tools come next, named about three times as often as LLM libraries.
• LangChain and LangGraph show up mostly at companies building AI features on top of models, not at the model labs. Vector database brands and no-code tools are almost absent.
• Some job ads name Claude Code, Cursor, Codex or MCP, often for product, design and marketing roles. They ask for the habit, not a specific licence.

How we counted the tools in AI job postings

We counted tool names in every active job on Tokn with two methods, and we give companies as well as jobs so one large recruiter can't skew the picture.

• Source: about 770 active jobs at 79 companies, as of 12 October 2026. You can browse all AI jobs on Tokn.
• Method 1: a dictionary of 54 tools and languages, matched against the full text of each posting. Most figures below come from here.
• Method 2: for tools outside that dictionary (Claude Code, Cursor, Codex, MCP, Hugging Face), a text search of the shortened summaries we store. These counts are minimums: Python shows up in 389 jobs with method 1 but only 182 with method 2.
• Caveats: Celonis has 31 active jobs, many near-identical across countries, so we give figures without Celonis where it matters. When a vendor names its own product (Anthropic, OpenAI, Snowflake), we leave those postings out or say so. Below 10 jobs, read a number as an example, not a statistic.

The short answer: the top 15 tools in AI job postings

The most requested tools are general engineering tools: Python, the three big clouds, Kubernetes, PyTorch and SQL. The first LLM-specific library, LangChain, only appears in 14th place, level with C++.

We left out React and Node.js, whose counts include false matches (the verb "react", "multi-node training"). RAG (70 jobs at 20 companies, 52 without Celonis) is a technique, so it sits with the LLM frameworks below.

We think this order is a sensible default for deciding what to learn: Python, one cloud and Kubernetes cover far more postings than LangChain plus a vector database.

In-demand AI tools by family

Grouped by family, programming languages lead, followed by infrastructure and cloud. Named LLM libraries appear in about 7% of jobs, vector databases and no-code tools in under 2% each.

Programming languages: Python, then SQL, TypeScript, Rust and Go

Python is asked for in 389 jobs at 61 companies: two-thirds of research and LLM jobs, three-quarters of data science jobs.

SQL comes next (109 jobs at 35 companies, 84 without Celonis). TypeScript (98 jobs at 33 companies) is named in 42% of LLM and GenAI jobs, often product engineering roles built on top of models. GitHub sees the same shift: TypeScript became the most-used language on GitHub by contributors in August 2025, overtaking Python and JavaScript.

Rust (65 jobs, 18 companies) and Go (64 jobs, 20 companies) are infrastructure languages. Rust demand comes largely from Lovable (14 jobs), Anthropic (8), Helsing (8) and Perplexity (6). Go appears in 30% of MLOps and infrastructure jobs.

ML frameworks: PyTorch leads, JAX is a lab skill

PyTorch is the default: 117 jobs at 36 companies (100 without Celonis), against 33 jobs at 22 companies for TensorFlow. In research roles the gap is wider, with PyTorch in 52% of jobs and TensorFlow in 11%.

JAX (37 jobs, 13 companies) is concentrated at model labs: Cohere and Mistral AI (9 jobs each), H Company (4) and Anthropic (3). It appears in 26% of research jobs. scikit-learn looks bigger than it is: 17 of its 31 jobs are Celonis postings, leaving 14 jobs elsewhere. Hugging Face appears in at least 3 postings outside its own.

LLM and agent frameworks: LangChain, LangGraph, LlamaIndex

LangChain is the most named LLM framework, with 42 jobs at 16 companies. Celonis supplies 14 and Elastic posts one role in 5 locations, so the real spread is narrower.

LangGraph, which LangChain describes as "an agent runtime and low-level orchestration framework", appears in 19 jobs at 9 companies. 11 of them are LLM and GenAI roles, where it is the most named framework (14% of jobs). LlamaIndex is rare: 6 jobs at 3 companies (Aircall, Artefact, Shift Technology). AutoGen and CrewAI appear in fewer than 5 postings each.

The bigger pattern is who asks: LangChain appears in about 1% of jobs at AI-native companies (8 jobs) against about 18% at other employers. Celonis inflates that second figure, but even without it the share is around 12%. Framework requests come mostly from companies building AI features on top of models (Elastic, GitLab, Aircall, Celonis), not from the labs. See the LLM and GenAI jobs to compare.

Developer usage puts LangChain first too: among developers who use an agent framework, Stack Overflow's 2026 survey found 30.2% use LangChain, 23.8% the OpenAI Agents SDK and 18.8% LangGraph (n = 4,400).

MCP: the protocol engineering ads name

MCP (Model Context Protocol) is named in at least 30 postings at 21 companies, 26 of them outside Anthropic. The official documentation calls it "an open-source standard for connecting AI applications to external systems". Anthropic introduced it in late 2024 and in December 2025 donated it to the Agentic AI Foundation, under the Linux Foundation.

Companies naming it include Elastic (one role in 5 locations), Supabase, Figma, Aircall, Doctolib and Mistral AI. More broadly, "agents" or "agentic" appear in 218 postings at 53 companies, close to 3 jobs in 10. If you target LLM and GenAI jobs, expect agent experience to come up.

Inference and GPU: CUDA, vLLM, SGLang

A small, specialised niche. CUDA appears in 25 jobs at 11 companies, and vLLM, "a fast and easy-to-use library for LLM inference and serving", in 14 jobs at 7 companies, led by H Company (4).

SGLang appears in at least 6 postings at 3 companies (H Company, ElevenLabs, Mistral AI), TensorRT in 2: examples rather than statistics, all inference roles at model companies.

Vector databases: the brand barely matters

Vector database brands are almost never named. Pinecone, Weaviate and pgvector each appear in one posting, the same one (Writer), and Qdrant in one (Dust). The concept matters more: "vector database", "vector store" or "vector search" appears in at least 13 postings at 10 companies, and RAG in 70. Learn retrieval and how to evaluate it, then pick up whichever store the team uses.

Cloud, infrastructure and MLOps

AWS leads the clouds (130 jobs, 40 companies), ahead of GCP (101 jobs, 32 companies) and Azure (89 jobs, 29 companies). Infrastructure roles often list all three, and Azure is relatively more common at employers that are not AI-native.

Kubernetes is the big surprise: 126 jobs at 38 companies, and 68% of MLOps and infrastructure jobs. Mistral AI (22 jobs), Anthropic (9), Perplexity (8) and Together AI (8) ask for it most. It appears in about 20% of jobs at AI-native companies against 7% elsewhere. Terraform follows (59 jobs, 23 companies; 38% of MLOps jobs), then Docker (32 jobs) and Ray (16 jobs, 7 companies).

MLOps platforms are rarely named: MLflow in 6 jobs at 3 companies, Weights & Biases and Kubeflow in one posting each.

Data tools: SQL first, then a warehouse

For data science jobs, SQL (69%) sits almost level with Python (75%). Next come BigQuery and dbt (20% each), then Databricks, Snowflake, GCP and Azure (16% each).

Databricks looks bigger than it is: 29 jobs, but only 6 companies and 14 Celonis postings. Snowflake (28 jobs, 12 companies) includes 7 of Snowflake's own. dbt is the most evenly spread: 17 jobs at 10 companies, none with more than 3.

AI coding assistants: Claude Code, Cursor, Codex, Copilot

Coding assistants do appear in job ads, in small numbers. Outside the vendors' own postings, we found Claude Code in at least 8 postings, Cursor in 9 and Codex in 7. Six of each are the same set of Celonis consulting roles; the rest come from Brevo, Granola and Scale AI. We did not find GitHub Copilot named in any posting summary.

Claude Code is "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools". Cursor presents itself as "your coding agent for building ambitious software". Codex is OpenAI's coding agent.

Usage is far ahead of what job ads spell out. Among AI-tool users, Stack Overflow's 2026 survey found 65.5% use Claude Code, 58.7% GitHub Copilot, 29.5% OpenAI Codex and 25.5% Cursor (n = 12,255). At work specifically, JetBrains' January 2026 AI Pulse put GitHub Copilot at 29% (40% at companies with 5,000+ employees), and Cursor and Claude Code at 18% each. The two surveys measure different populations.

The postings ask for a habit, not a licence. Brevo's AI Native Product Engineer ad names "Claude Code, Cursor, Codex, something equivalent", and asks for "real delegation of substantial work", not just generated unit tests. Black Forest Labs puts it bluntly: "Most of our code is written by agents."

No-code automation: n8n, Zapier, Make

No-code tools are rarely named. n8n appears in 6 jobs at 5 companies, all in AI ops and automation: Dust, Equativ, Artefact, Doctolib and Lovable. Zapier appears in 5 jobs at 4 companies (Doctrine, Dust, LangChain, Doctolib), mostly growth and operations roles.

Make is hard to measure: our dictionary only catches "make.com", and in summaries we found it only in two n8n postings naming competitors, so it is rarely named rather than absent. n8n and its peers help for AI ops and automation jobs, but they won't carry an application on their own.

"AI fluency" is a line in non-technical job ads too

The least obvious AI skill employers want is showing that AI is part of how you work. Brand-named assistants (Claude, Cursor, Claude Code, Codex) appear in about 18 postings outside the vendors, at 9 companies, mostly for product, design, consulting, SEO and growth roles. Broader phrasings like "AI-native" or "AI-assisted" add dozens more.

Some examples from current postings:

• Photoroom, Senior Product Manager (and three other PM roles): "You use Claude or equivalent every day, for prototyping, testing with users, or automating the parts of the job that used to fill your calendar."
• Synthesia, SEO Manager: "AI-native operator. You do the work in Claude or your AI tool of choice."
• Granola, Product Designer: prototyping "in Figma or code with AI tools like Cursor/Claude Code".
• Celonis, data and AI consulting roles: "Comfort working with AI-assisted development tools (e.g., Claude Code, Codex, Cursor, etc.). Ability to write clear specifications, evaluate AI-generated outputs".
• Doctrine, Growth Marketing: "this role runs 100% with AI, campaign production, copywriting, analysis, automation."

Read together, these ads ask for three things: daily use, a workflow you can explain, and judgment on what the model produces. Engineering ads more often name MCP and agents than a specific assistant.

Which AI skills employers want, by role

Each role family has its own short list. Research wants PyTorch and JAX, MLOps wants Kubernetes and Terraform, data science wants SQL, and product and marketing roles rarely name any tool at all.

For what each of these paths involves day to day, see our overview of the 8 AI career paths.

What the "best AI tools" lists get wrong

Most listicles overweight niche tools and underweight infrastructure.

Overrated:

• Vector database brands: Pinecone, Weaviate, Qdrant and pgvector appear in one posting each.
• No-code platforms: n8n and Zapier together cover 9 jobs.
• "OpenAI API" as a skill: 4 postings. Ads like Granola's ask for multi-provider experience instead.
• TensorFlow over PyTorch: PyTorch is named in more than three times as many jobs.

Underrated:

• Kubernetes and Terraform: 126 and 59 jobs.
• Rust and Go: 65 and 64 jobs, mostly infrastructure roles.
• SQL: 109 jobs, and 69% of data science roles.

AI tools to learn: an order that matches the postings

The right order depends on where you start. Here is what we would learn, based on what postings name most for each path.

Software engineer moving into AI

1. Python.
2. One cloud (AWS is the most named), then Docker and Kubernetes.
3. LLM APIs from more than one provider, RAG, and LangGraph or an equivalent.
4. MCP: build a small server for a tool you use.
5. Daily use of a coding agent, with work you can show.

Data scientist or analyst

1. Python and SQL, both solid.
2. dbt and one warehouse (BigQuery, Snowflake or Databricks).
3. PyTorch.
4. RAG and how to evaluate LLM outputs.

Research or inference engineer

1. PyTorch, then JAX.
2. CUDA and GPU performance.
3. vLLM or SGLang.
4. Kubernetes and Ray for distributed work.

Non-engineer (product, design, marketing, ops)

1. An assistant such as Claude or ChatGPT, daily, on real work.
2. Cursor or Claude Code to prototype.
3. One real automation in n8n or Zapier.
4. SQL basics, and Figma if you are in product or design.

Our step-by-step guide to getting a job in AI covers the rest of the process.

Find the jobs that ask for your stack

To check this against your own profile, read the postings. Start with all AI jobs on Tokn or a category page, and see which companies are hiring the most in Europe. Then create a job alert for your category and city, so new roles that name your tools reach you by email.

FAQ

What AI tools do job postings ask for most?

Python leads by far, in about half of AI jobs on Tokn as of 12 October 2026 (389 postings). Then come AWS and Kubernetes (about 17% each), PyTorch (15%), SQL (14%), GCP (13%) and Azure (12%). The most named LLM framework, LangChain, is in about 1 job in 18.

Do I need LangChain or LangGraph?

It depends on the employer. LangChain (42 jobs, 16 companies) and LangGraph (19 jobs, 9 companies) come mostly from companies building AI features on top of models, such as Elastic, GitLab and Aircall; model labs rarely name them. Learn the concepts (agents, tool calling, RAG) first, one framework second.

Do employers ask for vector databases like Pinecone or Weaviate?

Almost never by name: Pinecone, Weaviate, Qdrant and pgvector each appear in a single posting. Vector search as a concept and RAG (70 jobs) matter far more than any product.

Do job ads ask for ChatGPT, Claude or Copilot skills?

Some do, mostly for product, design, consulting and marketing roles, where ads name Claude, Cursor, Claude Code or Codex and ask for daily use. We did not find GitHub Copilot named in any posting summary. Engineering ads more often ask for MCP (at least 30 postings at 21 companies) and agent experience.

Are n8n, Zapier or Make worth learning for a job?

They help for AI ops, automation and growth roles, but are rarely named: n8n in 6 jobs, Zapier in 5, Make only in passing. Build one real automation you can explain, alongside Python or SQL.