v0.1.24

App questions now plan real MCP calls when your connector exposes the apps family, with per-tool costs measured. Which tools a connector exposes varies by plan and by server release; run /sw-config --show to see yours.

What's Changed

• Composition, comparability and client-register guards from a dogfooding review by @idan-yaron in #10
• Re-ground the apps family and guard against absence assertions by @idan-yaron in #11
• v0.1.24 by @idan-yaron in #12

Full Changelog: v0.1.23...v0.1.24

v0.1.23

What's Changed

• v0.1.23: full surface coverage, payload budget, cost-model corrections by @idan-yaron in #8

Full Changelog: v0.1.22...v0.1.23

Full Changelog: v0.1.22...v0.1.23

v0.1.22

What's Changed

• build: bump softprops/action-gh-release from 3.0.1 to 3.0.2 by @dependabot[bot] in #3
• build: bump actions/checkout from 7.0.0 to 7.0.1 by @dependabot[bot] in #4
• build: bump actions/setup-python from 6.3.0 to 7.0.0 by @dependabot[bot] in #5
• fix(sw-config): preserve the capability map during --refresh (v0.1.22) by @idan-yaron in #7

Full Changelog: v0.1.21...v0.1.22

v0.1.21

Re-grounds the plugin on the live Similarweb MCP surface, which grew from 80 tools to 113 and renamed 8 of them into a get-website-analysis-* namespace.

Fixed

• sw-channel-mix aborted on every run. Its required probe tool read as absent under the new names, and that tool's pinned absence outcome is ABORT.
• sw-competitive-teardown silently dropped its channel breakdown and PPC investment sections.
• sw-aeo-audit's AI-referral fallback called a tool that no longer exists.
• The Cowork teardown artifact failed on page load.
• 403 claims detection was silently broken. The server's 403 message changed wording and the "missing the required claims" phrase moved into the actions array, so message-gated detection would miss a real claims gap. Detection now keys on HTTP 403 plus error.code FORBIDDEN_ERROR only.

Added

• AI-referral measurement via the Gen AI traffic channel. get-website-analysis-traffic-channels now returns a ten-channel taxonomy including Gen AI with absolute visits at 10 credits per month, so sw-aeo-audit leads with it and costs less than the proxy it replaces. Naming the individual AI sources stays opt-in.
• Rename-transition safety. If a documented tool is absent but its retired predecessor is live, the plugin calls the predecessor and says so. Connectors do not all update together.
• A tool-name drift guard in build.py --validate, so the next rename wave fails the build instead of shipping. It scans every file type across the shipped trees, the generated mirrors, and the root docs.

Downloads

Platform	Asset
Cowork	similarweb-cowork-0.1.21.zip
Claude Code	similarweb-claude-code-0.1.21.zip
Codex	similarweb-codex-0.1.21.zip
Cursor	similarweb-cursor-0.1.21.zip

Install steps per platform are in docs/install-and-smoke-test.md.

What's Changed

• fix(mcp): re-ground on the 113-tool surface and restore the renamed tools (v0.1.21) by @idan-yaron in #6

Full Changelog: v0.1.20...v0.1.21

v0.1.20: M365 Copilot converter-input support

Groundwork for Microsoft 365 Copilot: the build now regenerates a converter-input companion for Microsoft's Claude-plugin conversion script on every release. See the README for what the plugin does.

In this release:

• New emit_m365 build target staging dist/m365/ as input for Microsoft's Convert-ClaudePluginToMOS3.ps1 (M365 Copilot Cowork): claude-code-shaped skills, a URL-only .mcp.json, and new M365 icons at the tree root. Built by build.py --build, source-build-only, NOT attached to releases and not installable as-is.
• Skill descriptions are scrubbed at M365 emit time of clauses that would misread on Microsoft's platform (where "Cowork" names Microsoft's product); the staged tree is LF-normalized and byte-stable across build hosts.
• Converter behavior grounded live (script SHA-256 335ab2bf, 2026-07-06): remote MCP servers configure by bare URL, skills copy byte-verbatim, PrivacyUrl and TermsOfUseUrl are required converter parameters.
• Validator additions: M365 icon dimension checks (PNG IHDR), description drop-table integrity, and a validate-time residual-Cowork gate.
• One clarity fix in sw-foundation-render's persistent-export note: names Anthropic Cowork explicitly and includes Microsoft Copilot among plain-markdown platforms.

Bundles: four platform zips below; install and smoke-test steps in docs/install-and-smoke-test.md.

v0.1.19

v0.1.19: cross-platform hardening

Makes the capability-map skills portable across shells and Python interpreters. See the README for what the plugin does and how to install.

In this release (since v0.1.18):

• The two capability-map skills (sw-config, sw-setup) no longer assume bash heredocs or a python3 binary. They now drive scripts/capmap.py subcommands with a --file option and a write-to-temp-file flow, so they work under PowerShell (Windows Codex) and where the interpreter is python rather than python3.
• capmap.py gains --file, sched, and recycle subcommands; stdin input stays back-compatible.
• The skill source is shared across bundles, so the same fix covers Claude Code and Cursor on Windows (Cowork runs on Linux; Claude.ai ships no foundations).

Bundles: the four platform zips are attached below. Per-environment install steps are in docs/install-and-smoke-test.md.

v0.1.18

v0.1.18: Codex bundle standardization

Brings the Codex bundle fully onto the Codex skill spec and verifies it end-to-end on a live Codex install. See the README for what the plugin does and how to install.

In this release (since v0.1.17):

• Per-skill agents/openai.yaml now uses the snake_case shape Codex's skill loader honors, so recipe install-card names and example prompts register correctly.
• The three foundation skills are no longer auto-selectable on Codex; recipes, the router, and the operators stay implicitly invocable.
• Corrected the Codex data-connection docs (codex mcp add or an [mcp_servers] block) and added the OpenAI-curated Similarweb connector as a one-click path, verified end-to-end with the recipes.
• build.py --validate now guards the openai.yaml shape and rejects descriptions that break strict-YAML skill loading on Codex (the cause of an older foundation-load failure).
• Documented a persistent-export path for non-Cowork platforms and surfaced the orchestrator sub-agents.

Bundles: the four platform zips are attached below. Per-environment install steps are in docs/install-and-smoke-test.md.

v0.1.17

Full Changelog: v0.1.15...v0.1.17

v0.1.15

similarweb v0.1.15

Presence-first handling of your account's tool surface, a working Cowork agent tier, and content fixes. Full docs and quickstart in the README.

In this release:

• Tool presence is now a first-class gating axis: detected live from the session's tool list at zero credits, with tools_absent and a tool_surface fingerprint in the capability map, per-recipe documented outcomes when a required tool is absent, a smoke retarget ladder, and a new module-not-exposed render pattern with a two-step remedy.
• Call-time unknown-tool detection is message-gated on live-grounded error wording, so a validation error is never mistaken for an absent tool.
• Cowork agents no longer declare a tools allowlist. Live verification showed the previous list did not block MCP calls; the removal makes the inherited tool surface explicit.
• The market-size brand-family table drops two incorrect entries (Figma under Adobe, Bose under Sony) and now states its maintenance posture.
• Stale labels fixed: the render output-targets line is version-neutral and the Codex install description reflects the 80-tool live surface.

Bundles attached: cowork, claude-code, codex, and cursor zips. Install and smoke-test steps: docs/install-and-smoke-test.md.

v0.1.14

Full-workspace QA release: recipes now dispatch as skills on every platform (fixing the Cowork "Unknown command" failure), and the grounding layer was re-validated against the live MCP. See the README for what the plugin is.

In this release:

• sw-router gains a Dispatch mechanics contract: recipes are applied as skills (exact listed name, inline-follow fallback), never invoked as slash commands; the Cowork hook hint, agents, and command bodies align with it.
• The repo root now carries .claude-plugin/marketplace.json, making the documented Claude Code /plugin marketplace add install path work.
• Hooks rewritten to the stdin-JSON contract: the Stop hook now functions, the prompt hint matches only the prompt field, and all hooks invoke via bash.
• New grounded assertion: get-websites-similar-sites-agg requires exactly a 3-month window when dates are passed explicitly; callers pin it.
• Smoke probes are bounded and reused (no double-charged target rank call, no accidental 36-month rank series), and the AEO window formula uses relative keywords.
• Honest cost guidance: re-priced router estimates, mandatory geography limit, and a consent gate on deduplicated-audience above 2 domains.
• Foundation skill descriptions are now strict-YAML-parser safe; two audience-overlap tier examples corrected; the Sources contract unified to a single line.
• Tool catalog re-grounded at the 80-tool live surface; validator adds marketplace, command, and em-dash checks, and CI fails on Codex mirror drift.

Bundles and per-platform install steps: docs/install-and-smoke-test.md.