Skip to content

competitive-landscape.md

Latest commit

 

History

History
221 lines (190 loc) · 13.5 KB

competitive-landscape.md

File metadata and controls

221 lines (190 loc) · 13.5 KB
Raw

Competitive Landscape & Long-Term Goals

Last updated: 2026-06-23

This document is the evidence base behind industry-leader-roadmap.md. The roadmap says what we will build; this file says who else is in the market, where the open lane is, and which long-term bets follow from that. Keep the two in sync: when the landscape shifts, update this file first, then re-derive the roadmap.

Star counts, install counts, and versions below are point-in-time reads from June 2026 and will drift. Treat them as relative signal, not live data.

North-star positioning

stata-code should be the most reliable, agent-native, typed way to run, inspect, repair, and audit Stata — winning on fidelity to the authoritative Stata runtime and referee-grade reproducibility, not on method count or raw breadth.

The single fact that defines our lane: no competitor ships all three of

  1. a typed error taxonomy (stable error.kind values an agent branches on, not return codes + red text it must string-match);
  2. typed r() / e() result contracts (not a generic results dump); and
  3. token-efficient by-reference artifacts for logs, graphs, and matrices.

Everything else is either AGPL/GPL-licensed, editor-bound, raw-log-only, or a Python/R reimplementation that does not touch the real Stata runtime.

The field (June 2026)

Tool License Exec Structured out Typed error kinds Frontends Maturity
stata-code (this) MIT pystata typed RunResult + r()/e() yes MCP + kernel + VS Code pre-1.0
tmonk/mcp-stata + workbench AGPL-3.0 pystata JSON + r()/e()/s() no (rc + red text) MCP + ext ~69★, very active
hanlulong/stata-mcp MIT pystata no (filtered raw log) no VS Code + MCP ~440★, ~15.6k installs
SepineTam/mcp-for-stata AGPL-3.0 do-file subprocess no (raw SMCL) no MCP + CLI (7+ agents) ~204★, very active
haoyu-haoyu/stata-ai-fusion MIT pexpect partial (r()/e()/c()) no (text-flagged) MCP + Skill + ext ~32★, new
stata_kernel (kylebarron) GPL-3.0 automation/console no (raw log) no Jupyter ~278★
nbstata (hugetim) GPL-3.0 pystata no (raw log + widgets) no Jupyter ~59★
pystata (StataCorp) proprietary native r()/e()→dict, mat→NumPy no (Py exceptions) IPython magics ships w/ Stata
StatsPAI (sibling, Python) MIT Python reimpl. yes (agent cards) partial (validation) MCP ~244★, daily

Reading the table

  • The genuine head-to-head is tmonk/mcp-stata — pystata-backed, MCP-native, returns r()/e(), ships a skills catalog, and has StataCorp-newsletter visibility. Its two gaps are our two wedges: (a) no typed error taxonomy (agents still parse return codes + preserved red text) and (b) AGPL-3.0, a hard blocker for commercial/embeddable adoption. MIT + typed kinds is the clean answer to both.
  • The adoption leader is hanlulong/stata-mcp (~15.6k installs). Despite using pystata — which could expose stored results — it surfaces filtered raw log text, not a schema. Distribution, not schema quality, is its moat; we must not assume a better contract auto-wins installs.
  • Jupyter kernels and pystata itself are human-facing, not agent-native: raw log + images, Python exceptions, no MCP, no by-reference economy.
  • The real category threat is the Python/R reimplementation wave (StatsPAI, rmcp): they took the "first agent-native econometrics" framing. We do not beat them on method count — we beat them by being the authoritative Stata leg they themselves reach for when cross-validating.

Why the error taxonomy is the defensible identity

A typed error taxonomy is the one capability no competitor has and the hardest to retrofit onto a raw-log design. It is also cheap for us to lead on because we already have the architecture (error.kind, error.suggestions, error.rc_label, pinpoint context). The 2026-06-23 core pass made this concrete:

  • audited RC_TO_KIND against StataCorp [P] error (Stata 19) and corrected multiple misclassifications (e.g. not_sorted is r(5), not the unrelated r(119)/r(459));
  • populated error.rc_label with Stata's canonical short message via label_for_rc() (it was silently empty for every real error before);
  • expanded suggestions_for() so nearly every common failure ships an actionable recovery hint.

The moat is not "we have error kinds" — it is "our error kinds are correct, labeled, and paired with a recovery action, verified against the manual." That is the kind of trust an empirical economist and a referee both need.

Long-term goals (6–12 months), by leverage

Ranked. Each ties back to a roadmap pillar and is phrased as a durable outcome, not a feature list. Status legend: ✅ shipped · 🟡 foundation shipped, stretch remaining · ⬜ not started.

  1. ✅ Own "agent-native typed Stata errors" as the headline. Shipped 2026-06-23. error.kind is a stable, manual-verified contract (audited against [P] error); every classified rc ships a canonical rc_label (label_for_rc) and, where actionable, a remediation suggestion. The agent recovery contract (error.recovery / recovery_for) gives a defined next action per kind: retry-as-is, change-code, or escalate. → Roadmap pillar 1 (reliable execution contract).
  2. ✅ Per-command typed r()/e() result contracts for mandated commands. Shipped 2026-06-23. RunResult.results.estimation is a typed coefficient table (term/b/se/statistic/p/CI + model_stats) from referee-grade r(table) (or e(b)/e(V) fallback). It now also carries command_family (ols/iv/gmm/panel/count/…) and command-aware diagnostics — the identification/spec tests economists must report (ivreg2/ivreghdfe weak-ID F + Hansen J, xtabond2 AR(2)/Hansen, reghdfe within-R²/absorbed FE, xtreg rho), surfaced only when present in e(). This is the StatsPAI-defense: referee-grade numbers from the exact mandated command. → Roadmap pillar 2.
  3. ✅ Reproducibility / provenance envelope. Shipped 2026-06-23. build_provenance() captures Stata version/edition, e(cmd), stata-code + schema versions, timestamp, seed, and per-package dependencies parsed from the script (ssc/net install → Provenance.packages); build_reproducible_do() renders a version-pinned, seed-set re-runnable .do; and build_submission_package() assembles a replication/journal bundle (do + PROVENANCE.json + README manifest). Stretch: package version pinning (vs. name only) and a journal-specific layout. → Roadmap pillar 1 + 3.
  4. ✅ Data-MCP integration bridge (FRED / World Bank / Census). Shipped 2026-06-23. verify_dataset() enforces the handoff's key checks (row/var counts, observation bounds, required columns) on the captured DatasetInfo — the executable companion to the data-mcp-handoff protocol, documented in references/structured-results.md. Stretch: first-class adapters that ferry source metadata (row hash, series ids) into the check automatically. No Stata MCP has done this composition. → Roadmap pillar 4.
  5. 🟡 Typed-schema-anchored skills catalog — replication audits, robustness sweeps, publication QA, legacy .do modernization — each anchored to typed results + provenance, under MIT. Foundation shipped: references/structured-results.md teaches agents to consume the typed contracts (results.estimation/diagnostics, error.recovery/rc_label, reproducible-do / submission bundles, verify_dataset). Stretch: the audit/robustness/QA recipe set in the skills lane (skills/stata-code/references/recipes/**). → Roadmap pillar 2.

Surface coverage & safety (shipped 2026-07-23)

The typed-contract lane (above) is the defensible identity; this cycle closed the most-cited surface and safety gaps against SepineTam/hanlulong so "covers every Stata usage scenario" is literally true, not aspirational.

  • ✅ Bash / plain-terminal surface. stata-code run (a .do file, -e snippets, or stdin → RunResult, text or --json, exit-coded) puts the full structured loop behind any agent that can shell out — Codex CLI, CI, a bare terminal — without requiring MCP. Routes through the same subprocess pool, so timeout/cancel/policy all apply. stata-code lint exposes the linter too.
  • ✅ One-command onboarding. stata-code setup --claude|--cursor|--vscode|--all writes the MCP server entry (merging, backing up, --dry-run), matching SepineTam's install --all. doctor stays read-only; setup is the opt-in mutating counterpart. Codex (TOML) / Claude Desktop are emitted as copy-paste snippets rather than edited in place.
  • ✅ Command-safety guard (core.policy, error.kind="policy_blocked"). Blocks shell / winexec / erase / rm / rmdir / ! before Stata runs, at both the pool boundary and the in-process runner; configurable via STATA_CODE_COMMAND_POLICY / _ALLOW / _BLOCK. Parity with SepineTam's 27-rule guard for the autonomous-overnight use case. A guard rail, not a sandbox.
  • ✅ Static pre-run lint (core.lint, MCP lint_do). Unbalanced braces, missing end, dangling /// — caught Stata-free before a run is spent. A token-economy play no competitor ships.

Closing the gaps vs. editor-first tools (shipped 2026-07-23)

Against stata-all-in-one (a polished, distributor-backed VS Code extension whose "AI Skill" hands agents raw log text), this cycle neutralized its two adoption advantages while keeping the typed-contract lane it cannot enter:

  • ✅ Console (batch) backend (core.console, Backend.CONSOLE). Drives the Stata CLI in batch mode and parses the log into the same typed RunResult (typed r()/e(), estimation matrices, the error taxonomy). Covers Stata 13+ with no pystata — removing their "we support 13+, you're pystata-only" edge, and unlike their AI Skill the output stays structured.
  • ✅ Zero-Python standalone binary (scripts/build_standalone.py + packaging/standalone.github-workflow.yml). With --backend console, a Python-free path to typed results — answering their "ready out of the box, no Python" pitch.
  • ✅ One-click VS Code onboarding. The extension provisions a workspace .venv + server on a single click / command, matching their frictionless install.
  • ✅ Reliability/token benchmark scaffold (benchmarks/) — the apparatus to turn "typed beats raw-log" into a published number.

Still open (next up), in priority order:

  1. ⬜ Streaming / progress for long runs. Real boottest / csdid jobs run for minutes; hanlulong streams, we batch. The schema already reserves log.complete:false. Highest-leverage remaining surface gap.
  2. ⬜ Hard timeout / cancel for the Jupyter kernel. The kernel still uses the in-process runner (no preemption). Move it onto the subprocess pool or an equivalent, resolving the documented interactivity-vs-safety tradeoff.
  3. ⬜ Human IDE polish to match stata-all-in-one where it also serves the agent loop: inline graph rendering + DPI export, a scalable data viewer, and an optional "attach to a running Stata" backend (shared live session).
  4. ⬜ Console backend depth — graph capture and persistent sessions; run the benchmark for real and publish the numbers.

Risks & threats

  • StataCorp native AI — LOW near-term, monitor. Stata 19 shipped classical H2O ML only; no LLM/copilot/agent feature is shipped or announced, and StataCorp frames AI as community-built. Watch New in Stata / StataNow for any shift from tutorials to a shipped feature. Our durable hedge: we require a genuine Stata license, so our incentives align with StataCorp's rather than competing with them.
  • Distribution gap. hanlulong already has ~15.6k installs and tmonk has newsletter visibility. A superior contract does not auto-win adoption; the typed-error / reproducibility story must reach economists where they are (SSC, Statalist, replication and referee communities).
  • Category framing already taken by Python/R. "Agent-native econometrics with structured results + MCP" converged across StatsPAI and rmcp. Defend by positioning on authoritative Stata-runtime fidelity, not breadth.
  • Generic code-exec substitution. A Jupyter-MCP + statsmodels sandbox can "do econometrics" with zero domain tooling. Defense: mandated Stata commands + verifiable typed e() contracts a generic sandbox cannot provide.
  • License contagion. The two structured/MCP-native competitors (tmonk, SepineTam) are AGPL; the kernels are GPL. To preserve our MIT clean-room wedge we must never vendor their code paths (see LICENSE-POLICY.md).
  • Naming/trademark. "Stata" is a StataCorp trademark and the *stata-mcp namespace is crowded. Keep the "not affiliated with StataCorp" disclaimer prominent and avoid implying endorsement.

Sources

Competitor repos and listings (June 2026): tmonk/mcp-stata, tmonk/stata-workbench; hanlulong/stata-mcp (VS Code Marketplace: DeepEcon.stata-mcp); SepineTam/mcp-for-stata (PyPI stata-mcp); haoyu-haoyu/stata-ai-fusion; kylebarron/stata_kernel; hugetim/nbstata; StataCorp pystata docs and New in Stata 19; brycewang-stanford/StatsPAI; finite-sample/rmcp; data MCPs (datacommonsorg/agent-toolkit, stefanoamorelli/fred-mcp-server, worldbank/data360-mcp). StataCorp [P] error (Stata 19, 2025) is the authoritative source for the error-code audit.