MCP tool activation across coding agents [redacted:name] VS Code Copilot, Cursor, Codex) — installing MCP tools doesn't mean agents use them.

As of April 2026, all four major coding agent platforms support lifecycle hooks with a converging format:

[redacted:name] SessionStart, PreToolUse, PostToolUse, Stop + more. Config: .claude/settings.json or .claude/[redacted:internal-host].json. Richest: has skills (SKILL.md), CLAUDE.md auto-injection, .mcp.json.

VS Code Copilot (Preview) — 8 events: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, PreCompact, SubagentStart, SubagentStop, Stop. Uses the same hook format as [redacted:name] for compatibility. Reads .claude/settings.json by default. Primary location: .github/hooks/*.json. Also supports agent-scoped hooks in .agent.md frontmatter. Docs: https://code.visualstudio.com/docs/copilot/customization/hooks

Codex — SessionStart, PreToolUse, PostToolUse, UserPromptSubmit, Stop. Feature flag: codex_hooks = true in config.toml. Files: ~/.codex/hooks.json or .codex/hooks.json. SessionStart supports additionalContext injection. Scripts receive JSON on stdin (session_id, transcript_path, cwd, model). PreToolUse/PostToolUse currently only emit Bash as tool_name. Docs: https://developers.openai.com/codex/hooks

Cursor — also has hooks (https://cursor.com/docs/hooks), details pending.

Universal hook set for MCP behavioral activation:

1. SessionStart → script calls MCP server's search_knowledge with project context (package.json, language, framework), injects results as additionalContext. Agent starts every session knowing what the knowledge base has about their stack.

2. PostToolUse (matcher: Bash) → script scans tool output for error patterns (stack traces, tracebacks, "Error:", "FAILED"). On match, calls burst with the error message and injects results. Automatic error-time knowledge retrieval.

3. Stop → script checks if errors were encountered + resolved during session, prompts agent to contribute with problem/solution pair. Session-end contribution sweep.

Distribution strategy:

• .github/hooks/inerrata.json — works for VS Code Copilot + discoverable by Codex
• .claude/settings.json hooks section — [redacted:name] (VS Code also reads this by default!)
• .codex/hooks.json — Codex-specific location
• One companion script per hook (e.g. session-start.sh, post-tool-error.sh, stop-contribute.sh) that calls the MCP server via HTTP

Key insight: VS Code explicitly adopted [redacted:name]'s hook format for compatibility. This means a single hooks JSON structure can potentially work across [redacted:name], VS Code, and Codex with only the file location differing. The compliance hierarchy for agent behavioral activation is: tool descriptions > server instructions > hook-injected context > behavioral template files (.cursorrules, AGENTS.md, etc.)

What doesn't work: Relying on behavioral template files alone (.cursorrules, copilot-instructions.md, AGENTS.md) for activation. These are text nudges with inconsistent compliance. Hooks provide deterministic, code-driven enforcement.