Datasette 1.0a43 arrived on November 15, 2026 with a security fix for private stored queries, streaming CSV and TSV exports, and a pile of 1.0 milestone polish—while companion plugins datasette-agent 0.6.0, datasette-apps 0.4.0, and datasette-maplibre 0.1a0 shipped the same weekend. Datasette.io's news post is the authoritative changelog; Simon Willison's release write-up adds context on Claude Fable 5.2 assisting the 1.0 push.
If you publish SQLite to the web for newsrooms, NGOs, or internal ops dashboards, this release is both patch Tuesday and feature Friday: close ACL leaks before agents get wider store-query permissions, then turn on streaming exports so million-row tables do not OOM your workers.
TL;DR #
| Item | Detail |
|---|---|
| Versions | **1.0a43** (1.0 alpha) and**0.65.6** (0.6 stable) |
| Security | Private stored-query metadata exposure fixed in 1.0a43;database download ACL fix in both lines |
| Advisory | Willison deemed the stored-query issue GitHub advisory–worthy (see theDatasette security advisories page ) |
| Feature flag | Response.stream() for output renderers; core CSV/TSV use it |
| Agent UX | datasette-agent 0.6 homepage prompt + table hash-mentions |
| Apps | datasette-apps 0.4 URL fragments for cross-app linking |
| Maps | datasette-maplibre 0.1a0 MapLibre + OpenFreeMap plugin |
| 1.0 progress | Willison: milestone 53% complete, heavyClaude Fable 5.2 use |
Security: private stored queries and database downloads #
Datasette.io summarizes two classes of bugs:
- Private stored queries — actors who should not see a private query could still readtitle and SQL text , even if they could not execute it. That is aconfidentiality breach for teams using stored queries as saved investigation playbooks.
- Database downloads — when a database is registered under aroute different from its filename , download ACLs could let denied actors fetch the file anyway.
Willison's personal release notes say one flaw deserved a published GitHub advisory. Multi-tenant instances—common in data journalism circles that adopted Datasette during COVID dashboards—should treat this as patch-now, not next sprint.
A protected key in a capsule evoking private stored-query metadata patched in Datasette 1.0a43
Upgrade matrix
| Track | Version | Action |
|---|---|---|
| Running 1.0 alphas | → 1.0a43 | pip install datasette==1.0a43 |
| Running 0.6x stable | → 0.65.6 | Same pin on 0.65.6 |
| Private queries unused | Still upgrade | Download ACL bug may still apply |
| Air-gapped single-user | Lower urgency | Metadata leak risk minimal |
After upgrading, audit view-query, store-query, and download-database permissions on each published database. The stored-query redesign from 1.0a31 made queries first-class rows; this patch closes a hole in that model.
Adapter pipes meeting through a junction evoking Datasette streaming CSV and TSV export renderers
Streaming exports and the 1.0 finishing list #
The headline feature is Response.stream() for output renderers. Built-in TSV and CSV exporters now stream row batches instead of buffering entire tables in memory—critical when reporters dump multi-gigabyte public-records tables.
Plugins can register as stream-capable, which matters for bespoke redacted exports in legal-tech stacks. Willison also shipped:
- Row JSON layout deemed ready for 1.0 freeze.
<datasette-copy-schema>web component for schema snippets.- Pretty-printed JSON columns on row pages.
DatasetteModal.confirm()JS helper.- View transitions CSS for smoother multi-page navigation (Safari/Chrome today; Firefox 162 in December 2026 per Willison).
He explicitly credits Claude Fable 5.2 for accelerating the 1.0 backlog—consistent with Fable-assisted creative coding posts elsewhere on explainx.ai.
Companion plugins from the same release train #
Willison bundled four ecosystem updates on November 15:
datasette-agent 0.6.0
Adds an agent prompt on the Datasette homepage with hash-mentions for tables and saved queries—basically steering LLM CLI tools toward live schema context. This pairs with the stateless MCP interest Willison flagged after mcp-explorer experiments.
datasette-apps 0.4.0
Sandboxed HTML/JS apps inside Datasette (Artifact-like, but SQL-aware) now read URL fragments, enabling one app to link another with structured state. Datasette Agent learns supported schemes so prompts can generate cross-app navigation.
datasette-maplibre 0.1a0
First new geospatial plugin in a while: MapLibre GL plus OpenFreeMap vector tiles, merging ideas from datasette-cluster-map and datasette-leaflet-geojson. Willison says Fable 5.2 scaffolded much of the repo in one prompt—a useful data point for geospatial agent demos teams comparing MapLibre to Google stacks.
datasette 1.0a43 proper
Treat the plugin releases as dependent features on the core security and streaming work—do not upgrade plugins without bumping core.
What developers should do this week #
- Pin versions in
requirements.txtor Docker images; rerun CI against 1.0a43. - Re-scan stored queries for sensitive SQL comments you assumed were private.
- Test streaming exports on your widest table; tune reverse-proxy timeouts.
- If you expose Datasette via MCP , re-run permission audits—agents amplifying metadata leaks is worse than browsers doing so.
- Read Willison's view transitions commit if you skin custom templates; animations respect
prefers-reduced-motion.
Concrete upgrade and test checklist
| Step | Why it matters |
|---|---|
Snapshot datasette.yaml and plugin pins |
Roll back if streaming plugins misbehave |
Run datasette --version in each environment |
Staging often lags production on SQLite publishers |
Hit /-/queries as a low-privilege test user |
Confirm private titles no longer visible cross-account |
| Attempt database download on renamed routes | Reproduces the ACL bug class patched in 0.65.6 |
| Export widest table to CSV/TSV while watching RSS | Validates streaming path under load |
| grep logs for agent plugins listing queries | Agents cache schema; flush after ACL changes |
Docker deployers should rebuild images rather than pip install inside running containers—Datasette's plugin ecosystem loads at process start, and orphaned workers are a common reason newsrooms think they patched when one pod did not.
Who gets hurt if you delay?
| Persona | Risk if you skip 1.0a43 |
|---|---|
| Investigative desk | Saved SQL in private queries leaks story angles |
| Multi-tenant SaaS | Customer A sees Customer B's named investigations |
| Agent automation | LLM tools exfiltrate query titles into thread logs |
| Public civic site | Lower risk on read-only single-tenant, but download bug may still apply |
Stored queries in the 1.0 permission model (context) #
Datasette's stored query overhaul landed in 1.0a31: canned queries became database rows with private defaults, trusted YAML imports, and granular store-query / execute-write-sql permissions. That design is why the November 2026 bug stung—private queries are first-class secrets, not ephemeral browser state.
If you migrated from 0.64.x, revisit:
- Which queries are
is_trusted: truein YAML imports. - Whether
execute-write-sqlshould be narrower after agent plugins arrive. - How
datasette-agenthomepage prompts expose table names via hash-mentions (feature in 0.6.0).
Willison's stack is a reference architecture for small-team AI publishing: Fable drafts plugins, Datasette enforces ACLs, agents read only what permissions allow. A metadata leak breaks that chain even when execution stays blocked.
Streaming architecture notes for operators #
Before Response.stream(), exporters often materialized whole tables—fine for thousand-row demos, fatal for FOIA dumps. Streaming shifts failure modes:
- Reverse proxies may buffer unless chunked encoding is honored.
- Plugins registering renderers must declare stream capability explicitly.
- Downstream notebooks that assumed finite JSON payloads need pagination updates.
If you previously cron-exported CSV to S3, measure time-to-first-byte after upgrading; it should drop on wide tables. If it does not, check for plugins still using non-streaming render paths.
Map plugin and apps: should you install them now? #
datasette-maplibre 0.1a0 is alpha—fine for newsroom maps mirroring MapLibre geospatial demos you already run in static HTML, but expect breaking URL schema changes.
datasette-apps 0.4.0 matters when you want Artifact-like internal tools without shipping a second React app. Fragment linking is the enabling detail for multi-app newsroom kits (overview app → record app).
Neither plugin fixes security; upgrade core first, then bump plugins in a second deploy so you can separate regression signals.
Willison's 1.0 milestone cadence #
Willison wrote that the 1.0 milestone is 53% complete, with Claude Fable 5.2 accelerating UI polish and bug-backlog burns. That matters for dependency planning: alpha releases will keep shipping quickly, and security patches may land between feature blog posts you bookmarked weeks ago.
Teams standardizing on Datasette for election or investigative dashboards should subscribe to Datasette.io news and Simon Willison's feed, not only GitHub releases—advisories sometimes appear in narrative posts before release tags propagate to PyPI mirrors.
If you maintain a fork, diff not only Python but also default templates and static assets; view-transition CSS arrived as a small commit with outsized UX impact.
Relationship to the LLM CLI and MCP experiments
Willison's llm CLI and mcp-explorer tooling sit adjacent to Datasette in his stack. When datasette-agent puts prompts on the homepage, you are looking at the same philosophy as stateless MCP: constrain what agents can touch, log the surface, and keep humans one click away from raw SQL.
That is why ACL patches belong in the same incident ticket as agent upgrades—your threat model is now curious journalist + curious model, not only malicious outsider.
FAQ-style gaps people search #
Does the stored-query bug let attackers run SQL?
Willison's note: exposure covered title and SQL text, not execution rights. Still toxic if queries encode PII filters, law-enforcement joins, or unreleased story angles.
Is 1.0 finally stable?
No—this is 1.0a43, alpha. Production teams on 0.6x should stay on 0.65.6 until Willison tags 1.0 stable.
How does this compare to chaining MCP + SQLite elsewhere?
Datasette's value is opinionated read-only publishing with ACLs. The patch shows why hand-rolled FastAPI + SQLite tools need the same query-catalog ACL tests Datasette just learned.
Should I agent demos during the upgrade?
Yes, if agents can list stored queries or echo schema metadata in threads. Bring agents back only after you complete the private-query visibility test and confirm logs did not already capture leaked titles during the vulnerable window. If you demoed a sensitive investigation this week, rotate any named query that lived in a private catalog.
For public demos, pin datasette==1.0a43 in the Dockerfile you ship to conferences; unpinned pip install datasette pulled vulnerable alphas for anyone who cloned your repo after 1.0a42.
Related reading #
Version numbers and security scope match Datasette.io and Willison's posts as of November 16, 2026 UTC.