agentl

Website · Getting started · Connectors · PyPI · Roadmap · Discussions

Agentic ETL in YAML. Read files, APIs and databases, let an LLM classify or extract each record against a schema, check the data, and deliver it. Per-record cache, budgets, retries and a live dashboard come built in.

• Pay once per record: a re-run of unchanged data makes no model calls, and a job that failed halfway only pays for what did not finish (before and after).
• Data you can trust: check steps with fixed rules and an agent review before delivery.
• Works where your data is: CSV, Excel, PDF, SQL, Postgres, S3, SharePoint, Google Sheets, Kafka, SQS, webhooks, and an MCP server for Claude Code and Cursor.
• Honest numbers: benchmarks and evals against plain Python, including where it loses.

Status: beta (v0.9). Minor releases may still change the API; every change is in the changelog.

For AI agents: the docs are at agentl.com.br/llms.txt, and every docs page has a raw Markdown copy (<page>/index.md).

Install

pip install agentl-etl                 # core: CSV, JSON, HTTP, SQLite, runner, dashboard
pip install 'agentl-etl[office,pdf]'   # Excel, Word, PDF
pip install 'agentl-etl[anthropic]'    # Claude as the LLM provider
pip install 'agentl-etl[all]'

The distribution is named agentl-etl on PyPI; the import and the CLI are both agentl.

A pipeline

name: invoices
llm:
  provider: anthropic          # or "echo" to run offline
  model: claude-opus-5-5

steps:
  - id: invoices
    use: csv.read
    with: { path: data/invoices.csv }

  - id: customers
    use: json.read
    with: { path: data/customers.json }

  - id: matched
    use: join
    input: [invoices, customers]
    with: { left: invoices, right: customers, on: tax_id, how: left }

  - id: classify
    agent: Classify the expense category and flag unknown clients for review.
    input: matched
    output: { number: str, customer: str?, category: str?, needs_review: bool? }

  - id: save
    use: sqlite.write
    input: classify
    with: { db: out/invoices.db, table: invoices }

notify:
  channels:
    slack: { webhook: "${SLACK_WEBHOOK}" }
  rules:
    - on: run.failed
      to: [slack]
      summary: agent             # the LLM explains the failure in 3 lines
    - on: item.rejected
      when: "rate > 0.05"
      to: [slack]

agentl validate pipeline.yaml
agentl run pipeline.yaml --show classify
agentl diagram pipeline.yaml      # Mermaid
agentl dashboard                  # http://127.0.0.1:7070

You can also write extract:, transform: and load: sections instead of steps:. Inside transform and load, a step without input reads from the previous one.

Concepts

Everything is a step. A step gets records (list[dict]) from its inputs and returns records. Extract, transform and load are labels, not different machinery. Steps form a graph, so ETL, ELT, fan-out and joins all fit.

Agent steps run once per record, in parallel (concurrency), and must declare an output schema (str, int, float, bool, list[str], dict; a trailing ? means optional). Outputs are validated; a record that fails twice is rejected, not the whole run, and shows up as an item.rejected event. Results are cached per record in .agentl/agentl.db, so re-running after a crash only pays for what is left.

Secrets come from the environment: ${VAR} or ${VAR:-default}. Never put them in the YAML.

Built-in connectors

file.read csv.read csv.write json.read json.write txt.read excel.read parquet.read pdf.read docx.read doc.read xml.read html.read http.get http.post sqlite.read sqlite.write select filter dedupe join concat

file.read detects the format and expands ZIP/GZ archives. Readers accept path (glob), url (HTTP download) or upstream records. agentl connectors lists what is installed. http.get accepts {field} placeholders in the URL to make one request per upstream record, plus items: data.results to pick the list out of the response.

Extending

from agentl import connector, register_provider


@connector("crm.contacts")
def contacts(input, since=None):
    return [{"id": 1, "email": "a@b.com"}]

Packages can register connectors through the agentl.connectors entry point group, so pip install agentl-yourthing is enough. LLM providers plug in with register_provider(name, factory); the factory returns an object with structured(instruction, record, schema) and text(instruction, content).

Events

The runner emits run.started, step.started, step.succeeded, step.failed, item.rejected, source.empty, run.succeeded and run.failed. They feed the store, the dashboard and notification rules (slack, webhook, console).

Roadmap

Next up: a pandas bridge, Postgres with upsert and docs for AI agents. See the full public roadmap and vote on the issues you need.

Contributing

Bug reports, connector requests and pull requests are welcome. Read CONTRIBUTING.md for the setup and the checks every PR must pass, and SECURITY.md to report vulnerabilities privately. Release history is in CHANGELOG.md.

Sponsors

agentl is free and MIT licensed. If it saves you time, consider sponsoring the project.

License

MIT