Skip to main content

ts-agents docs

ts-agents is a general-purpose, CLI-first time-series toolkit for external agents and automation. Agents discover methods, run analyses on their own data files, and get inspectable artifacts with explicit quality signals across task families:

Task family Start with
First look, data quality, periodicity workflow run inspect-series
Anomalies and unusual events tool run find_discords --input ...
Change-points and regimes tool run segment_changepoint --input ...
Spectra, frequencies, coherence tool run compute_psd --input ... --time-col t
Trend and seasonality tool run stl_decompose --input ...
Activity recognition workflow run activity-recognition
Classification tool run compare_classifiers
Forecasting workflow run forecast-series / forecast-panel

Project-specific preparation, custom evaluation, and comprehensive reporting can still require code. These docs capture the project’s durable design choices so interfaces can evolve without rewriting the core toolchain.

Why Autonomous Agents Like It

  • Bootstrap with one call: ts-agents capabilities --json returns the full agent-facing bootstrap surface (workflows, tools, sandboxes, install_profile, status-contract guidance).
  • Find tools by task: tool search anomaly --json matches task words and synonyms, and every tool takes data with tool run <tool> --input <file>.
  • Discover before acting: workflow show --json / tool show --json expose cli_templates, source_options, global_options, default_output_behavior, required extras, and availability in the current environment.
  • Strict machine envelopes: every --json response carries schema_version: "1.0", strict JSON (no raw NaN), and typed exit codes for validation / dependency / permission / timeout failures.
  • Quality signals: workflow payloads expose result.status (ok / degraded) and result.data.quality_flags, while the CLI envelope exposes top-level quality_status, degraded, and requires_review.
  • Run lifecycle metadata: generated run IDs, run_manifest.json, and non-clobbering default output directories with --overwrite / --resume semantics — so long multi-turn sessions remain traceable and safe to rerun.
  • Sandbox parity + explicit fallback: local, subprocess, docker, daytona, and modal all stage workflow artifacts back to the host output directory; sandbox doctor probes readiness; the executor never silently downgrades isolation.

Get Started

  1. Quickstart — install and run your first workflow in under a minute
  2. CLI Examples — the CLI contract by example, across task families
  3. Walkthroughs — step-by-step guides for the first-class workflows
  4. Skills Guide — task-family skills for external agents
  5. Panel and foundation-model forecasting — global GBM/NHITS models and Darts zero-shot checkpoints on long-format panels

Learn More

  • Evaluation — benchmark evidence and rerun instructions
  • Design — why the project is structured around a stable CLI contract for agents
  • Distribution — packaging, release, and publishing workflow

Talks