Python, not bash. Large data stays in a cache as handles, never in the prompt.
Every run is logged — and eval-backed.

---

Most data-agent tooling makes you pick between giving a model a shell (unsafe, irreproducible) and single-shot code-gen (no state, no multi-step). data-harness is the controlled middle path: the model works through a constrained Python interpreter, large objects live in a SessionCache and are exposed as compact handle snapshots — so a 100k-row table never hits the context window — every turn is logged to JSONL, and a built-in evaluation harness measures quality and cost across providers.

Principles

• Python, not bash — one controlled execution surface: no shell side-effects, no destructive commands, reproducible runs.
• Handles, not payloads — large data lives in the cache; only snapshots reach the model, so context (and cost) stay flat as data grows.
• Measured, not vibes — a first-class eval harness with programmatic graders, multi-turn cases, cost, and tracked leaderboards.

Features

• One-liner — ask(df, "...") in Python, or dh "..." data.csv from the shell.
• Charts & SQL — automatic matplotlib capture; a DuckDB / SQLAlchemy sql_query tool.
• Many providers, one key — OpenAI, Anthropic, DeepSeek, Qwen, Google, Z.ai… via OpenRouter.
• MCP bridge — connect any MCP server (Postgres, SQLite, filesystem…) and use its tools, with progressive disclosure + handle/snapshot.
• Production controls — subprocess sandbox, an approval gate, and a zero-token replay cache.
• Evaluation — bespoke / hard / large-data suites + WikiTableQuestions, with multi-turn cases, cost, and JSON-tracked results.
• Composable — ask/Chat over Agent over Harness; async + streaming; subagents; progressive connectors.

---

Install

pip install data-harness          # core
pip install "data-harness[all]"   # + openai, charts, duckdb, sqlalchemy, notebook, eval

Pick individual extras as needed: [openai], [viz], [duckdb], [sql], [notebook], [eval]. Requires Python 3.10+.

---

Quickstart

Ask a question about a DataFrame in one line. ask() resolves a provider from your environment, loads the data into the session cache, runs the agent, and returns a RunResult:

import pandas as pd
from data_harness import ask

df = pd.read_csv("sales.csv")
result = ask(df, "What was total revenue, and which month was highest?")

print(result.text)      # the written answer
print(result.value)     # the structured result the model computed via answer()
result.charts           # any charts it rendered (notebook-friendly)

Reach many providers through one key with OpenRouter — a provider/model id auto-routes there. Set OPENROUTER_API_KEY:

ask(df, "plot revenue by month", model="deepseek/deepseek-v4-flash")
ask(df, "summarise the data",   model="google/gemini-2.5-flash-lite")
ask(df, "which region grew fastest?", model="qwen/qwen3.5-flash-02-23")

Without OpenRouter, ask() falls back to ANTHROPIC_API_KEY / OPENAI_API_KEY / DEEPSEEK_API_KEY. In a notebook, the returned RunResult renders prose, the value, and charts inline (there's also a %%ask magic via %load_ext data_harness.notebook).

---

Command line & demo

Ask from the shell with dh (also installed as data-harness) — point at one or more files, or pipe CSV via stdin:

dh "What was total revenue?" sales.csv
dh "Join these and find the top region" orders.csv customers.csv
cat sales.csv | dh "median order amount" --json

A Streamlit demo app (pip install "data-harness[demo]"):

uv run streamlit run examples/streamlit_app.py

Multi-turn chat

from data_harness import Chat

chat = Chat(df)
chat.ask("What was total revenue?")
chat.ask("Which month was highest?")   # remembers context (shared cache + history)

---

Charts & SQL

matplotlib runs inside the interpreter; open figures are captured automatically as artefacts — the image bytes live on disk and never enter the message history or logs (only a path does):

result = ask(df, "Plot revenue by region as a bar chart.")
result.charts[0]        # a ChartArtifact; renders inline in Jupyter

With DuckDB installed, ask exposes a sql_query tool over your DataFrames; point it at a real database with a SQLAlchemy URL:

ask(df, "Use SQL to get total revenue per region.")          # DuckDB, in-process

from data_harness import Agent
agent = Agent.from_dataframe(df).enable_sql(engine_url="postgresql://...")
agent.run("Top 5 customers by spend last quarter?")

---

Production controls

from data_harness import Agent, ExecutionCache

agent = Agent.from_dataframe(df).enable_cache(ExecutionCache("cache.json"))  # 0-token replays
sandboxed = Agent.from_dataframe(df, execution="subprocess")                 # isolated process
gated = Agent.from_dataframe(df, on_code=lambda code: (print(code), True)[1]) # approve code
preview = Agent.from_dataframe(df, code_only=True)                            # dry-run, never executes

• Code-replay cache — a repeat question over the same data schema replays the recorded code with no model call (zero turns, zero tokens), and stays correct when the data changes.
• Subprocess sandbox — interpreter code runs in a separate process with networking disabled and CPU/wall-clock limits; handles cross by value, results merge back.
• Approval gate — on_code sees every code block before execution and can block it; code_only=True returns the code without running it.

---

Evaluation

A first-class harness to measure how well an agent answers real data questions — across models, with programmatic grading that leans on the structured .value (no LLM judge needed for most cases).

from data_harness.eval import evaluate_matrix, fetch_openrouter_prices, hard_suite

models = ["deepseek/deepseek-v4-flash", "qwen/qwen3.5-flash-02-23",
          "openai/gpt-5-nano", "google/gemini-2.5-flash-lite"]
report = evaluate_matrix(hard_suite(), models)
print(report.to_markdown(fetch_openrouter_prices(models)))  # accuracy / turns / tokens / cost

• Suites — bespoke_suite() (smoke), hard_suite() (multi-table joins, deep multi-step, stateful multi-turn), large_data_suite() (100k-row frames answerable only via the handle, with a snapshot trap), and load_wikitablequestions() (public table-QA, the model differentiator).
• Case types — single-shot EvalCase and multi-turn ConversationCase (graded turns over one Chat session, testing SessionCache persistence).
• Graders — numeric, contains, exact, dataframe_equals, chart_produced, refuses, all_of/any_of.
• Reporting — leaderboards with per-model cost, per-category breakdowns, and to_dict()/to_json() for results tracked in evals/results/.

Results are committed as readable leaderboards — see evals/results/SUMMARY.md (a table per suite: accuracy, turns, tokens, cost).

What the runs show: the structured/large/stateful suites saturate at ~100% across recent models — i.e. the design is robust (even small, cheap models handle 100k-row data via the handle for ~$0.002 and pass the snapshot trap). Model differentiation shows up on messy real-world data — WikiTableQuestions spreads recent models 64%→96%. See the Evaluation guide.

---

Lower-level Agent and Harness

ask/Chat are conveniences over Agent, itself a thin layer over Harness. Drop down for full control:

from data_harness import Agent
from data_harness.providers.anthropic import AnthropicAdapter

agent = Agent(adapter=AnthropicAdapter(model="claude-sonnet-4-6"), system="You are a data analyst.")
print(agent.run("Compute the mean of [1, 2, 3, 4, 5]."))

Component	Role
Harness	The ReAct loop — messages, tool dispatch, reminders, JSONL logging
SessionCache	Handle-based store; keeps large objects out of message history
ProviderAdapter	Translates provider SDK responses into harness types
python_interpreter	The model's only execution surface
ConnectorRegistry	Hides connector tools until the model loads them
Subagent	Isolated worker with explicit state transfer

Async + streaming (AsyncAgent.run_stream), progressive connectors, and subagents are all supported — see examples/advanced_wiring.py and the docs.

---

Examples & tests

uv run python examples/live_demo.py          # ask()/charts/SQL on a cheap model
uv run python examples/eval_demo.py --suite hard   # multi-model eval leaderboard (cost)
uv run python examples/cache_benchmark.py    # replay-cache benchmark (no API key)
uv run python -m pytest tests/ -m "not live"       # offline test suite

examples/demo.ipynb is an executed end-to-end notebook.

---

Sandbox disclaimer

The in-process interpreter uses AST checks and restricted globals to reduce accidental misuse — it is not a container sandbox. For stronger isolation use execution="subprocess" (separate process, no network, resource limits). Neither is hardened for untrusted input.

---

Links

• Docs: https://maxkskhor.github.io/data-harness/
• Changelog: CHANGELOG.md
• Design series: three-part write-up
• License: MIT