feat(mcp): split OAuth endpoints and restrict directory tools

.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
     {
       "name": "synter",
       "source": "./",
-      "description": "Connect every ad platform, build audiences from signals, generate on-brand creative, launch cross-platform campaigns, and reallocate spend by ROAS — one conversation, every platform. You direct, the agents execute."
+      "description": "Plan and operate advertising campaigns through Synter’s hosted MCP server over OAuth. Review performance, write ad copy, and approve campaign changes before execution."
     }
   ]
 }

.claude-plugin/plugin.json

@@ -2,8 +2,8 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "synter",
   "displayName": "Synter",
-  "description": "The AI Agent Operator for Ads. Connect every ad platform, build audiences from signals and first-party data, generate on-brand creative, launch cross-platform campaigns, and reallocate spend by ROAS — all in one conversation. You direct, the agents execute. Nothing ships without your approval.",
-  "version": "1.1.2",
+  "description": "Plan and operate advertising campaigns through Synter’s hosted MCP server over OAuth. Review performance, write ad copy, and approve campaign changes before execution.",
+  "version": "1.2.0",
   "license": "MIT",
   "author": {
     "name": "Synter",
@@ -25,5 +25,17 @@
     "linkedin-ads",
     "mcp",
     "synter"
-  ]
+  ],
+  "hooks": {
+    "SessionStart": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "\"${CLAUDE_PLUGIN_ROOT}\"/scripts/session-context.sh"
+          }
+        ]
+      }
+    ]
+  }
 }

.cursor-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "Synter's official plugin marketplace. The AI Agent Operator for Ads.",
-    "version": "1.1.2"
+    "version": "1.2.0"
   },
   "plugins": [
     {
       "name": "synter",
       "source": ".",
       "logo": "assets/logo.png",
-      "description": "The AI Agent Operator for Ads. Connect 21 ad platforms, build audiences, generate on-brand creative, launch cross-platform campaigns, and reallocate spend by ROAS — one conversation, every platform. You direct, the agents execute. Nothing that spends money ships without your approval."
+      "description": "Plan and operate advertising campaigns through Synter’s hosted MCP server over OAuth. Review performance, write ad copy, and approve campaign changes before execution."
     }
   ]
 }

.cursor-plugin/plugin.json

@@ -2,8 +2,8 @@
   "$schema": "https://cursor.com/schemas/cursor-plugin/plugin.json",
   "name": "synter",
   "displayName": "Synter",
-  "version": "1.1.2",
-  "description": "The AI Agent Operator for Ads. Connect 21 ad platforms, build audiences, generate on-brand creative, launch cross-platform campaigns, and reallocate spend by ROAS — one conversation, every platform. You direct, the agents execute. Nothing that spends money ships without your approval.",
+  "version": "1.2.0",
+  "description": "Plan and operate advertising campaigns through Synter’s hosted MCP server over OAuth. Review performance, write ad copy, and approve campaign changes before execution.",
   "author": {
     "name": "Synter",
     "email": "joel@synterai.com"
@@ -29,15 +29,5 @@
   "skills": "./skills/",
   "agents": "./agents/",
   "rules": "./rules/",
-  "mcpServers": "./mcp.json",
-  "variables": {
-    "type": "object",
-    "properties": {
-      "SYNTER_API_KEY": {
-        "type": "string",
-        "title": "Synter API key",
-        "description": "Optional for onboarding. Paste your syn_... key from syntermedia.ai/developer to use connected ad accounts. Leave blank to run /synter:quickstart and the GA4 tools."
-      }
-    }
-  }
+  "mcpServers": "./mcp.json"
 }

.github/workflows/directory-tools.yml

@@ -0,0 +1,23 @@
+name: Directory tools
+
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+
+permissions:
+  contents: read
+
+jobs:
+  directory-tools:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+      - name: Check instruction tool references
+        run: node scripts/check-directory-tools.js
+      - name: Test repository validators
+        run: node --test scripts/*.test.js

.github/workflows/release.yml

@@ -24,6 +24,12 @@ jobs:
       - name: Run manifest validation
         run: node scripts/validate-manifests.js
 
+      - name: Validate directory tool references
+        run: node scripts/check-directory-tools.js
+
+      - name: Test repository validators
+        run: node --test scripts/*.test.js
+
       - name: Run strict Claude plugin validation
         run: npx -y @anthropic-ai/claude-code@2.1.263 plugin validate . --strict
 

.github/workflows/validate.yml

@@ -2,8 +2,6 @@ name: Validate
 
 on:
   pull_request:
-    branches:
-      - main
 
 jobs:
   validate-manifests:

CHANGELOG.md

@@ -2,6 +2,13 @@
 
 All notable changes to the Synter plugin are documented here. This project follows [semantic versioning](https://semver.org).
 
+## [1.2.0] - 2026-10-02
+
+- Cursor connects to `https://mcp.synterai.com`; Claude retains `https://mcp.syntermedia.ai` to match the connector under review. Both use browser OAuth.
+- Setup and support documentation now describe the hosted connection, shipped skills, and local hook accurately.
+- Removed unsupported workflows and limited instruction tool references to the approved 34-tool directory. Ad-copy writing and existing-asset selection remain supported.
+- Added offline directory validation and an optional OAuth-authenticated live MCP comparison.
+
 ## [1.1.2] — 2026-09-06
 
 ### Fixed
@@ -26,22 +33,16 @@ All notable changes to the Synter plugin are documented here. This project follo
 ## [1.0.0] — 2026-08-27
 
 ### Changed
-- **Approval gating is universal again.** The Solo/Scale tier carve-out that allowed direct single-turn execution has been reverted — every tier now requires explicit approval before any create/enable/budget/launch/pause call that spends money. Reads (`list_*`, `get_*`, `pull_*`, `verify_*`, reports) remain free to run without asking. The session-start hook, output style, and brand-and-safety context all describe the same universal gate again.
+- Account mutations and spend changes require explicit user approval. Read-only reports and account lookups run without asking.
 - Version canonicalized to **1.0.0** across `.claude-plugin/plugin.json`, `.cursor-plugin/plugin.json`, and `.cursor-plugin/marketplace.json`, closing the skew between the Claude and Cursor packaging surfaces.
 
 ### Added
 - **Packaging and release automation** — a zip build produces `dist/synter-plugin-<version>.zip` with the plugin at the archive root, matching Claude Desktop's zip-upload layout. CI validates every PR (manifest JSON, marketplace/plugin name match, skill frontmatter, `.mcp.json`/`mcp.json` interpolation guards) and a release workflow attaches the built zip to the GitHub Release on each `synter--v*` tag push.
 
-### Fixed
-- README no longer tells Claude Code installers they can "leave it blank" — `userConfig.synter_api_key` is `required: true` and a blank key yields zero Synter tools (see 0.1.1 below). The install steps now point straight at [syntermedia.ai/developer](https://syntermedia.ai/developer) and clarify that `.mcp.json` (Claude) and `mcp.json` (Cursor) use different key-interpolation syntax and must not be cross-wired.
-
 ## [0.4.1] — 2026-08-19
 
 ### Added
-- **Cursor Plugin packaging** — `.cursor-plugin/plugin.json`, `.cursor-plugin/marketplace.json`, and `mcp.json` so the same repo can be imported as a Cursor Team Marketplace and submitted to the [Cursor Marketplace](https://cursor.com/marketplace/publish). The hosted Synter MCP (`https://mcp.syntermedia.ai`) takes `SYNTER_API_KEY` via Cursor plugin variables. Listing metadata includes the Synter mark (`assets/logo.png`, 1024×1024) and the marketplace description on both the plugin manifest and the marketplace index.
-
-### Fixed
-- Cursor install no longer requires `SYNTER_API_KEY` up front (onboarding and free GA4 tools work without one). HTTP MCP declares `type: "http"`. Logo uses an absolute GitHub URL so the marketplace crawler does not 404. Cursor hooks use `sessionStart` instead of Claude's `SessionStart`.
+- Cursor plugin manifest, marketplace index, and remote MCP configuration.
 
 ## [0.4.0] — 2026-08-04
 
@@ -60,7 +61,6 @@ Fourteen new skills deepening diagnostics, measurement, and creative coverage 
 - **Platform depth** — `meta-ads-diagnostics` (Learning Phase, fatigue, overlap, Advantage+), `google-ads-quality-score`, `google-shopping-optimizer` (feeds, disapprovals, Shopping vs PMax), `performance-max-optimizer` (asset groups, brand cannibalization), `linkedin-ads-targeting` (ABM, lead gen forms).
 - **Measurement & data** — `pixel-capi-auditor` (dedup, Event Match Quality), `first-party-data-strategy` (Enhanced Conversions, server-side tagging, Consent Mode v2), `incrementality-testing` (geo holdouts, lift studies), `attribution` companions via `measure_incrementality` and `get_attribution`.
 - **Strategy** — `retargeting-sequence-designer` (staged funnels, recency windows, exclusions), `seasonal-budget-planner` (build your own seasonality index from your account data), `audience-expansion-strategy` (lookalike ladders, seed quality, layering), `geo-integrity-audit` (location-criteria and click-geography verification).
-- **Creative** — `video-ad-scriptwriter` (platform-specific scripts with timing marks), `ugc-creative-brief` (briefs, hooks, usage-rights structure).
 
 ## [0.2.0] — 2026-07-22
 
@@ -73,33 +73,21 @@ Broadened skill and agent coverage so the operator can plan, port, and run campa
 - Agents `media-buyer` and `campaign-strategist` now point at the `replicate` workflow and the platform playbooks.
 
 ### Changed
-- **`campaign-preflight`** rewritten to use only public MCP tools. The prior version referenced internal backend scripts (`*_validate_campaign_structure`, `run_gaql`, `verify_website_tracking`, `POST /tools/run` blocks) that the public MCP does not expose, and carried a hardcoded country blocklist. Now grounded in `run_gaql_query`, `ga4_list_conversions`, `get_gtm_tag`, and `verify_pixel_ownership`.
-- **`mmm-budget-planner`** rewritten around the public `optimize_budget`, `get_attribution`, and `measure_incrementality` tools instead of internal `mmm_*` training scripts that are not available through the MCP.
+- **`campaign-preflight`** moved from internal backend scripts to public MCP checks.
+- **`mmm-budget-planner`** rewritten around the public `optimize_budget`, `get_attribution`, and `measure_incrementality` tools instead of internal scripts that are not available through the MCP.
 
 ### Note on exclusions
 The new `kill-scale-rules`, `platform-benchmarks`, and `pre-pause-analysis` skills ship **generic methodology only**. Synter's proprietary numeric thresholds and cost-benchmark tables remain withheld — these skills teach the approach and defer to the account's own data, they do not carry the internal numbers. Internal runbooks and infrastructure-coupled skills are still not shipped.
 
-## 0.1.1 — 2026-07-20
-
-### Fixed
-- **API key is now required at plugin enable time.** The previous config invited users to
-  "leave blank to onboard," but the hosted Synter MCP server (mcp.syntermedia.ai) rejects
-  unauthenticated connections with a 401 — so a blank key meant the MCP server silently
-  failed to load and Claude had skill files but zero Synter tools. Requiring the key up
-  front prevents the broken skills-without-tools state. Get a key at syntermedia.ai/developer.
-
-
 ## [0.1.0] — 2026-06-16
 
 Initial release. The AI Agent Operator for Ads, packaged for Claude Code / Claude Desktop.
 
 ### Added
-- **MCP server** — Synter Advertising Platform (`https://mcp.syntermedia.ai`), cross-platform ad read/write. Free GA4 + onboarding tools work without a key.
-- **Workflow skills** (`/synter:*`) — quickstart, connect, audience, creative, launch, optimize, report, help.
-- **Craft skills** — 9 generic advertising-methodology skills: ad-policy-compliance, anomaly-detector, campaign-preflight, campaign-structure-auditor, creative-fatigue-detector, executive-reporting, landing-page-creation, mmm-budget-planner, roas-calculator. Deliberately excluded from the public package: internal-execution runbooks (Doppler/DB-coupled), Synter's proprietary operating playbooks (kill/scale thresholds, cost-benchmark + budget-split logic), and skills referencing internal infrastructure (pixel pipeline, table names, non-public tools).
+- **Workflow skills** (`/synter:*`) — quickstart, connect, audience, launch, optimize, report, help.
+- **Craft skills** — 9 generic advertising-methodology skills: ad-policy-compliance, anomaly-detector, campaign-preflight, campaign-structure-auditor, creative-fatigue-detector, executive-reporting, mmm-budget-planner, roas-calculator. Deliberately excluded from the public package: internal-execution runbooks (Doppler/DB-coupled), Synter's proprietary operating playbooks (kill/scale thresholds, cost-benchmark + budget-split logic), and skills referencing internal infrastructure (pixel pipeline, table names, non-public tools).
 - **Agents** — campaign-strategist, media-buyer, audience-builder, creative-director, budget-optimizer, performance-analyst.
 - **Hook** — SessionStart primer enforcing the money-safety rules and brand voice.
 - **Output style** — `synter` agent voice.
 - **Context** — bundled brand-voice and safety rules referenced by skills and agents.
-- `userConfig.synter_api_key` for turnkey, secure key entry at enable time.
 - **SDK runner** (`sdk/synter-agent.mjs`) — headless [`@anthropic-ai/claude-agent-sdk`](https://www.npmjs.com/package/@anthropic-ai/claude-agent-sdk) harness that loads the plugin and runs the operator unattended. Read-only by default (spend/mutation auto-denied via `canUseTool`); `--allow-writes` to lift. Verified: SDK loads all 23 skills + 6 agents.

CLAUDE_COMMUNITY_SUBMISSION.md

@@ -30,7 +30,7 @@ complementary catalogs. Do not treat connector review as plugin publication.
 | Category | Marketing and advertising |
 | Short description | Operate cross-channel advertising with approval-gated Synter agents, skills, and MCP tools. |
 | Brand icon | `assets/logo.png` — canonical transparent 1024×1024 Synter mark |
-| Support | `https://syntermedia.ai/contact` |
+| Support | `https://syntermedia.ai/support` |
 | Privacy policy | `https://syntermedia.ai/privacy` |
 | Terms | `https://syntermedia.ai/terms` |
 
@@ -91,11 +91,8 @@ Direct installation is not public-directory discovery.
 - [x] Production OAuth protected-resource and authorization-server metadata
   return `200`; a protected tool call returns `401` with a
   `WWW-Authenticate` resource-metadata challenge.
-- [x] The production MCP tool surface has 195 tools; every tool supplies a
-  title plus boolean `readOnlyHint` and `destructiveHint` annotations.
-- [x] Clean Claude Code configuration can add `Synter-Media-AI/plugin`, install
-  `synter@synter`, and inventory all 57 skills, 7 agents, 1 hook, and 1 MCP
-  server.
+- [ ] Run the optional live directory check with reviewer-authorized OAuth before submission; the checked-in directory contains 34 tools.
+- [ ] Re-inventory the updated package: 52 skills, 6 agents, one Claude session hook, and one remote MCP server per client.
 - [ ] Provider reviewer exercises `/synter:quickstart`, one read-only report,
   and one write preview without approving the write.
 - [ ] Provider reviewer uses a Synter-owned test workspace with sample data;
@@ -107,9 +104,8 @@ Direct installation is not public-directory discovery.
 
 ## Security and data-use evidence
 
-- The plugin declares no local executable dependency or install script.
-- The only remote MCP destination is `https://mcp.syntermedia.ai`, which uses
-  OAuth 2.1 with PKCE for Claude clients.
+- Claude invokes a Bash session hook that prints static guidance through `cat`. Cursor declares no hook. The repository also contains a manually invoked Node SDK runner and development scripts; no compiled binaries are bundled.
+- Claude uses `https://mcp.syntermedia.ai` to match the connector under review. Cursor separately uses `https://mcp.synterai.com`. Both plugin configurations use browser OAuth without static credentials or headers.
 - The session hook prints a static operating and approval-safety primer into
   session context; it does not read or transmit credentials.
 - Advertising writes are approval-gated. The bundled guidance does not turn a
@@ -138,14 +134,13 @@ boundary before any external spend-changing action.
 | --- | --- |
 | Public source and license | Public GitHub repository; MIT `LICENSE` |
 | OAuth for authenticated remote MCP | OAuth metadata, DCR, PKCE, and protected-tool `401` challenge at `https://mcp.syntermedia.ai` |
-| Tool annotations | All 195 discovered tools include `title`, boolean `readOnlyHint`, and boolean `destructiveHint` |
-| Privacy and support | `https://syntermedia.ai/privacy`; `https://syntermedia.ai/contact` |
+| Tool directory | 34 checked-in tool names; live annotations require a fresh authenticated review |
+| Privacy and support | `https://syntermedia.ai/privacy`; `https://syntermedia.ai/support` |
 | Narrow, visible instructions | Human-readable skills and agents in this repository; no encoded or remotely loaded behavioral instructions |
 | Approval boundary | Spend-changing operations require explicit approval; a model response is not approval |
 | Test account | Must be supplied privately by Synter in the submission form |
 
-The plugin manages advertising workflows and can invoke creative tools, so the
-submission must describe those capabilities accurately. Anthropic retains
+The plugin manages advertising workflows, writes text ads, and selects existing assets. The submission must describe those capabilities accurately. Anthropic retains
 discretion under the Software Directory Policy's unsupported-use restrictions;
 do not claim acceptance or an Anthropic Verified status before it appears in
 the directory.