v0.10.1 — useful first call and repeat-use SDKs

What changed

• Makes prediction-market preflight the first Agent interaction across MCP prompts and install paths.
• Fixes false homepage classification and bounds official-document search output.
• Preserves a privacy-safe installation identity across supported Agent clients.
• Adds TypeScript and Python risk middleware with shadow, warn, and enforce modes.
• Adds bounded outcome reporting and first-10-Agent pilot measurement.

Install

MCP bridge:

npx -y @mmvv1638/404-directory-mcp@0.10.1 --source your-agent

TypeScript risk SDK v0.1.0:

npm install https://github.com/MM-sheng/404-directory/releases/download/v0.10.1/mmvv1638-404-directory-risk-sdk-0.1.0.tgz

Python risk SDK v0.1.0 (Python 3.11+):

python -m pip install https://github.com/MM-sheng/404-directory/releases/download/v0.10.1/directory404_risk-0.1.0-py3-none-any.whl

Both SDK artifacts expose SHA-256 digests in the GitHub Release asset metadata and passed clean-environment import checks. Production at https://404.directory serves v0.10.1.

404.directory v0.10.0 — prediction-market risk preflight

Summary

Production release adding Polymarket settlement and execution-risk preflight for AI Agents.

• MCP: evaluate_prediction_market, report_prediction_market_outcome
• MCP: evaluate_tool_risk, report_tool_outcome (from v0.9.4 merge)
• REST: /v1/prediction-markets/evaluations/*
• 16 MCP tools on production (was 12)
• Database migrations 0007/0008 applied to Neon production

Verification

• Production: https://404.directory/health → 0.10.0
• Cloud Run revision: directory-404-v010-postmigrate
• Rollback: directory-404-aidisc-cbda6a7 (tag aidisc-canary, 0% traffic)

Not included

Does not predict outcomes, place orders, access wallets, or custody funds.

404.directory v0.9.3 — first useful calls for AI Agents

404.directory v0.9.3 turns installation into a task-oriented Agent workflow.

What changed

• Three MCP Prompts route Agents into real tasks: research-official-docs, verify-public-deployment, and evaluate-agent-tool.
• OpenAI Responses API clients can use a copy-ready remote MCP request that forces one useful tool call.
• Tool-only clients receive privacy-safe installation identity through a dedicated non-personal bearer token.
• The activation funnel now separates prompt discovery, prompt rendering, tool attempts, failures, and successful executions.
• Connection pages are no-store so generated installation identities are never shared through caches.

Install

npx -y @mmvv1638/404-directory-mcp@0.9.3 --source github-release

For Cursor, Claude, Codex, VS Code, OpenAI Responses API, and generic MCP configuration, use:

https://404.directory/connect?source=github-release

Complete one useful task

Ask your Agent to do something you already need:

• search current OpenAI, Microsoft Learn, AWS, or Cloudflare documentation and cite first-party sources;
• verify a public deployment against an expected status or release string; or
• discover and compare trusted read-only MCP tools for a real workflow.

Installation and probes do not count as adoption. The public target counts only de-duplicated external Agents that receive at least one non-error tool result.

• Live metric: https://404.directory/v1/metrics/agents
• Activation funnel: https://404.directory/v1/metrics/activation
• Privacy: https://404.directory/privacy
• Production health: https://404.directory/health

404.directory v0.9.2

Agent-ready v0.9.2

• Publishes the identity-preserving @mmvv1638/404-directory-mcp stdio bridge for Codex, Claude Code, Cursor, Cline, VS Code/Copilot, and compatible MCP clients.
• Keeps one stable random non-personal Agent ID per client while production stores only an HMAC digest after activity.
• Exposes exactly 12 read-only MCP tools for official documentation search, public deployment verification, and trust-aware tool discovery.
• Adds privacy-safe activation, reliability, retention, source, client, tool, provider, latency, and canonical-error measurement.
• Publishes matching metadata to npm and the Official MCP Registry.

Install

npx -y @mmvv1638/404-directory-mcp@0.9.2 --source your-source

No account or API key is required. Setup: https://404.directory/connect?source=github-release

Verification

A clean public-package acceptance run completed MCP initialize, listed exactly 12 tools, and returned a non-error result from search_tools. Production smoke passed verify_web, search_official_docs, and search_tools. Internal verification is excluded from the external-Agent adoption metric.

404.directory v0.9.1 — universal Agent installer

What changed

• Added a dependency-free universal stdio MCP bridge with one stable privacy-safe Agent ID per installation.
• Added the bridge to the human and Agent-readable connection guides with source attribution.
• Synchronized the service, Claude/Codex plugin manifests, and Official MCP Registry metadata.
• Production exposes 12 executable read-only tools for official documentation search, web verification, tool discovery, trust inspection, and controlled invocation.

Install

{
  "mcpServers": {
    "404-directory": {
      "command": "npx",
      "args": ["-y", "@mmvv1638/404-directory-mcp", "--source", "github-release"]
    }
  }
}

No account or API key is required. After installation, ask the Agent to complete one useful call such as search_official_docs. Only a de-duplicated external Agent with a successful tool execution counts toward adoption; probes and project self-tests are excluded.

Setup: https://404.directory/connect?source=github-release
Live metric: https://404.directory/v1/metrics/agents
Glama connector: https://glama.ai/mcp/connectors/io.github.MM-sheng/404-directory

404.directory v0.9.0

Human Agent activation flow for real external adoption.

External Agent pilot

Join the first-10 real external Agent activation pilot:

#1

Only genuine user tasks with a non-error tool result qualify. Probes, owner tests, anonymous calls, and duplicates are excluded.

Connect a real Agent

Use the one-minute installer for Cursor, VS Code / GitHub Copilot, Claude Code, or Codex:

https://404.directory/connect?source=github-release

After installing, ask your Agent:

Use search_official_docs to find the current official guidance for MCP Streamable HTTP. Cite the first-party sources and distinguish facts from inference.

A connection counts only after a non-error tool result. Probes, internal tests, anonymous calls, and duplicate IDs remain excluded.

• Live adoption target: https://404.directory/v1/metrics/agents
• Privacy-safe activation funnel: https://404.directory/v1/metrics/activation
• Agent-readable setup: https://404.directory/connect.md?source=github-release
• Documentation: https://404.directory/docs

Included in v0.9.0

• Browser-ready /connect flow with one-click Cursor and VS Code/Copilot installation.
• Claude Code plugin/direct MCP and Codex configuration.
• Privacy-safe per-install Agent identity and source attribution.
• First useful successful tool call guidance.
• Agent Plugin, Claude Plugin, Codex Plugin, and Agent Skill packages.

Validation: 87 tests pass, 1 is intentionally skipped; TypeScript and Drizzle migration checks pass.

v0.8.6

Publishes a directly installable lightweight Claude marketplace package. The isolated 36 KiB plugin contains only the validated Skill, native Claude manifest, dependency-free identity bridge, and license, avoiding installation of the application development dependency tree.

v0.8.5

Adds a native Claude Code and Cowork plugin manifest plus a Claude-specific identity-preserving MCP bridge configuration. Passes the official Claude plugin validator while retaining Agent Plugins 1.0, OpenAI/Codex plugin, Agent Skill, and MCP Registry compatibility.

v0.8.4

Adds an identity-preserving Agent Plugins bridge. Each installation generates one stable random local Agent ID under PLUGIN_DATA and sends it only to the fixed 404.directory MCP endpoint; the service stores only an HMAC digest after successful tool execution. This lets real plugin usage contribute to privacy-safe deduplicated adoption metrics.

v0.8.3

Adds portable Agent Plugins 1.0 packaging with root plugin.json and mcp.json, preserving existing OpenAI/Codex and MCP Registry compatibility. Includes schema validation tests and deploys the same 12-tool hosted read-only MCP service.