See Fluid Forge in action

Five short demos. Each one starts with a problem and ends with a number from the recording. They are terminal recordings rendered as SVG: some were captured from a live run, some were scripted against the bundled examples/ fixtures. Click play, read the takeaway banner, move on.

Recorded against CLI 0.10.0

The recordings were made with CLI 0.10.0 and have not been re-recorded for 0.18.1. Command names and flags in them still exist in 0.18.1 (the commands shown were checked against fluid --help), but the output, timings and costs on screen are what the recording showed, not what 0.18.1 prints today. For current output, follow Get Started.

Pick your problem

Your problem	Watch this	Time	What the recording shows
Writing a data contract per source by hand	`$0.03` per data product	~50 s	Three LLM providers, one flag swapped, the same contract shape.
Setting up acquisition for one Postgres source	Six months → sixty seconds	~40 s	A Bronze contract from a Postgres URL, run on embedded DuckDB.
An interview that asks what it could infer	23 questions, skipped	~50 s	A 47 ms scan answers most of the interview, four questions are asked.
A pipeline broke and you need the cause	Skip the panic	~75 s	`fluid runs` narrows a failure to a fix and `fluid ship` redeploys it.
An agent loop that accumulates tool history	`$0.50` → `$0.05`	~40 s	One env var selects the compaction strategy.

$0.03 per data product

fluid forge data-model turns eight lines of intent YAML into a contract. The recording runs the same intent through three providers (Anthropic Haiku 4.5, OpenAI gpt-4.1-mini, local Ollama gemma2) and shows the token counts and costs it measured. You choose the provider with --llm-provider.

fluid forge data-model — same intent, three providers, real cost figures

Eight lines of intent YAML. Anthropic, OpenAI, Ollama — same flag swapped each time. The recording shows all three producing a valid contract and the same dbt project layout, with the token counts and costs of that run.

Pairs with Forge Data Model and LLM Providers. Long-form animated reel preserved at /forge_docs/reels/forge-in-action.html.

Six months → sixty seconds

fluid init --discover postgres://… introspects a Postgres source and writes a Bronze acquisition contract. The recording validates it and applies it against embedded DuckDB. The contract's engine: field selects the runner (duckdb, dlt, meltano, airbyte, kafka-connect and debezium are build runners in the CLI), so moving off embedded mode is a contract edit; whether a given engine accepts a given source is a question for that engine's page.

fluid init --discover postgres://... — Bronze contract in 60 seconds

Connect, scan information_schema (47 ms), infer 28 tables / 143 columns / 12 PII candidates / 8 foreign keys, emit a complete Bronze contract.fluid.yaml, validate, apply against embedded DuckDB. 6.2 seconds total. Then show the engine swap path.

Pairs with Source-Aligned Acquisition, Postgres → DuckDB walkthrough, and fluid init --discover. Long-form animated reel preserved at /forge_docs/reels/source-aligned-bronze.html.

23 questions, skipped

fluid forge scans the working directory before it asks anything. In the recording the scan takes 47 ms, offers a 5-mode picker, asks four questions (most accept an inferred default), shows cost-cap progress as it runs, and shows a pre-write panel listing what will and will not change.

fluid forge — guided UX in action

47 ms welcome scan finds 3 CSVs + 2 dbt models + 1 README, infers domain (finance) and PII (5 columns). 5-mode picker. 4 questions answered (most accept the inferred default with ↵). Cost-cap progress in real time ($0.000 → $0.021 of $0.050 cap). Pre-write panel shows exactly what will + won't change.

Pairs with Guided fluid forge UX and the fluid forge reference. Long-form animated reel preserved at /forge_docs/reels/guided-forge-ux.html.

Skip the panic

A freshness SLA breaches overnight. The recording uses fluid runs status to see the failing runs, fluid runs logs --component dlq for the cause, fluid runs diff for what changed since the last good run, a one-line contract fix, and fluid ship to redeploy.

3am Slack ping → ship in 90 seconds

PagerDuty: freshness SLA breached. fluid runs status shows 3 consecutive failures. fluid runs logs --component dlq surfaces the root cause (CHECK constraint NOT NULL, 47 partial-window customers). fluid runs diff shows what changed since the last OK run. One-line contract fix: NOT_NULL → NOT_NULL_WHERE customer_age_days >= 30. fluid ship. 87 seconds end-to-end. Recovered 12,361 rows.

Pairs with fluid runs, fluid retention, fluid secrets, fluid stats, and Typed CLI Errors. Long-form animated reel preserved at /forge_docs/reels/day2-ops.html.

$0.50 → $0.05

A long agent loop accumulates tool results, and each turn carries the previous ones, so cost grows faster than the number of turns. FLUID_COMPACTION_STRATEGY selects how the loop compacts: truncate, summarize (LLM-backed) or hybrid (cheap path first, summariser as a safety net). The recording compares them on a 20-turn run.

Agent-loop compaction — three strategies, real before/after costs

20-turn baseline: $0.503/run, super-linear context bloat (5K → 67K → 298K tokens). Three strategies side-by-side: truncate (5.8× cheaper), summarize (9.3× cheaper), hybrid (10.5× cheaper). One env var: FLUID_COMPACTION_STRATEGY=hybrid. No code changes.

Pairs with Agentic primitives → Token-budget pre-flight & compaction. Long-form animated reel preserved at /forge_docs/reels/compaction-and-warnings.html.

More demos in 30 seconds each

The full library covers AWS, a Snowflake live-auth dry-run, agent-policy enforcement, blank scaffolds, policy compilation, and the local quickstart:

→ Browse all demos

How these are sourced

Each cast above is a deterministic asciinema-rendered SVG generated by a Python pipeline (scripts/cast_builder.py → scrub-cast.py → svg-term). The narrative lives in scripts/demos/<name>.py — clone, edit, regenerate. Each cast plays inline in the browser without GIFs, a JS framework or third-party CDNs.

Token counts, durations, and costs in the LLM-driven recordings (forge-multi-provider, agent-compaction) were captured from runs on the dates they were recorded; prices change, so treat them as that run's figures. The latency numbers in the other recordings (source-aligned-bronze, guided-forge-ux, day2-ops) come from the v0.10.0 stack against the bundled examples/ fixtures.

The original animated HTML reels are preserved at /reels/ for anyone who wants the longer-form pacing — each cast section above links through to its corresponding reel.