Telexed

AI news through a solo-operator lens — only what changes your day7 of 412

FILTER[All][Agents & tools][Models & API][Generative media][Infra & SaaS][ASO & growth][Indie business][Idea signals][Other][★6+ high-signal]

Simon Willison ✕clear filters

Fri, May 151 dispatches

#0412
#0412Other Simon Willison5 days ago
`datasette-llm-limits` `0.1a0` adds per-user LLM spend caps
40radar
DatasetteOpen-source data publishing tool — strong plugin extensibility
LLM usage inside Datasette can now be hard-capped by user or globally with rolling windows and USD limits. Useful if you expose prompts to users; otherwise the impact is narrow but the cost-control pattern is worth copying.
- Works with datasette-llm and datasette-llm-accountant, so enforcement sits on top of existing usage tracking.
- Config supports per-user or global limits, with windows like rolling-24h and explicit amount_usd caps.
- Example policy sets $1.00 per actor over 24 hours, which is a clean template for metered AI features.
- Best fit is multi-user internal tools or lightweight SaaS surfaces where prompt access needs a hard budget guardrail.
Source: simonwillison.net/2026/May/15/datasette-llm-limits/#atomRead original →
40radar
PHOTO
FIG-4121:1

Thu, May 141 dispatches

#0411
#0411Infra & SaaS Simon Willison6 days ago
`datasette-ip-rate-limit` `0.1a0`, path-level IP throttling for crawler abuse
50radar
datasette-ip-rate-limitDatasette plugin — path-level IP throttling rules
Bad crawlers hitting hot paths can now be blocked with 60 req/60s per-rule limits. Path-specific throttling in datasette.yml makes this a practical low-cost ops guardrail.
- Production config reads Fly-Client-IP and sets max_keys to 10000. It fits proxy-based deployments without extra app logic.
- Exemptions for /static/* and /-/turnstile* keep assets and bot checks flowing while narrowing protection to expensive routes.
- The example rule targets /global-power-plants/* and /legislators/*: 60 requests per 60s, then 20s blocked.
- Still an alpha at 0.1a0, but the config surface is small enough to copy into lightweight self-hosted services quickly.
Source: simonwillison.net/2026/May/14/datasette-ip-rate-limit/#aRead original →
50radar
PHOTO
FIG-4111:1

Wed, May 133 dispatches

#0410
#0410Other Simon Willison6 days ago
`CSP Allow-list Experiment`: Interactive `fetch()` Recovery Inside Sandboxed `iframe`s
40radar
A sandboxed iframe can catch blocked fetch() attempts, ask the parent to whitelist the origin, then reload with updated connect-src. Useful pattern for shipping safer user-script sandboxes without hardcoding every API upfront.
- Runs apps under default-src 'none' in a sandboxed iframe, then escalates blocked network origins to the parent for approval.
- Blocked fetch() calls surface the exact failed origin, so users can add only that domain to connect-src instead of loosening CSP broadly.
- The flow includes prompt, allow-list update, and refresh, turning CSP failures into a usable permission UX rather than a dead end.
- Good fit for code playgrounds, HTML preview tools, and user-generated app builders where external API access is needed but must stay constrained.
Source: simonwillison.net/2026/May/13/csp-allow/#atom-everythingRead original →
FIG-4101:1
40radar
FIG-4101:1
#0409
#0409Other Simon Willison7 days ago
`datasette` `1.0a29` fixes a nasty test race and smooths zero-row/mobile UX
40radar
DatasetteData tool — web app for exploring and publishing SQLite
A thread-safety bug that could segfault tests is now fixed, alongside small but practical UI cleanup for empty tables and Mobile Safari. Useful if you embed datasette in admin tools or data products, but not a broad market-moving release.
- Datasette.close() had a race with in-flight threaded queries and could trigger a segfault in tests; that class of bug is expensive to debug once it hits CI.
- Empty tables now still show headers and column options, so schema-first browsing works even with 0 rows.
- Mobile Safari column action dialog rendering was fixed, removing one of the more annoying iPhone/iPad admin UI papercuts.
- A new TokenRestrictions.abbreviated(datasette) helper generates "_r" dictionaries, which matters mainly for plugin and internal API users.
Source: simonwillison.net/2026/May/12/datasette/#atom-everythingRead original →
40radar
PHOTO
FIG-4091:1
#0408
#0408Agents & tools Simon Willison7 days ago
`llm` `0.32a2` adds `responses` API support for OpenAI reasoning models
70radar
llmCLI tool — exposes reasoning and tool-call flow
OpenAI reasoning models now route through /v1/responses, so llm can show summarized reasoning tokens alongside tool-use runs. Useful if you script GPT-5-class workflows from the CLI and want better visibility without building your own trace UI.
- OpenAI reasoning-capable models in llm now use /v1/responses instead of chat/completions, matching the newer tool-calling path.
- Reasoning summaries are surfaced directly in CLI output with separate coloring, which makes agent runs easier to inspect during debugging.
- -R and --hide-reasoning let you suppress that trace when logs need to stay clean or machine-readable.
Source: simonwillison.net/2026/May/12/llm/#atom-everythingRead original →
70radar
PHOTO
FIG-4081:1

Tue, May 121 dispatches

#0407
#0407Agents & tools Simon Willisonlast week
Run `llm` Directly from a Script Shebang
50radar
LLMCLI tool — turns prompt files into executable scripts
Plain-text prompts and YAML templates can become runnable scripts with llm. Useful for tiny automations and tool-calling helpers when shell-level speed matters.
- A shebang like #!/usr/bin/env -S llm -f turns an English text file into an executable prompt, lowering friction for one-off generators.
- Adding -T enables tool calls inside the script, so time lookups or other helpers can run without wrapping everything in Python.
- YAML templates can define Python functions and execute them via llm -t, pushing lightweight agent behavior into a single file.
- The shown calc example chains multiply and add with --td debug output, making tool execution transparent enough for quick internal utilities.
Source: simonwillison.net/2026/May/11/llm-shebang/#atom-everythiRead original →
50radar
PHOTO
FIG-4071:1

Sat, May 91 dispatches

#0406
#0406Agents & tools Simon Willison2 weeks ago
`Claude Code`: Ask for HTML, Not Markdown
40radar
Claude CodeCoding agent — HTML outputs make long reviews easier
HTML outputs unlock SVG, navigation, and interactive explanations instead of plain prose. Token efficiency matters less now, so rich review artifacts and code walkthroughs are worth trying immediately.
- A concrete prompt asks for an HTML PR review with inline diff annotations and severity colors, turning review output into a browsable artifact.
- The older reason to prefer Markdown was the 8,192-token era; that tradeoff weakened as context windows grew.
- HTML lets the model embed SVG, navigation, widgets, and styling in one response, which improves long technical explanations.
- A test run with gpt-5.5 on an obfuscated Linux exploit produced a usable interactive explainer, though prompt focus still mattered.
Source: simonwillison.net/2026/May/8/unreasonable-effectiveness-Read original →
FIG-4061:1
40radar
FIG-4061:1

`datasette-llm-limits` `0.1a0` adds per-user LLM spend caps

`datasette-ip-rate-limit` `0.1a0`, path-level IP throttling for crawler abuse

`CSP Allow-list Experiment`: Interactive `fetch()` Recovery Inside Sandboxed `iframe`s

`datasette` `1.0a29` fixes a nasty test race and smooths zero-row/mobile UX

`llm` `0.32a2` adds `responses` API support for OpenAI reasoning models

Run `llm` Directly from a Script Shebang

`Claude Code`: Ask for HTML, Not Markdown