Skip to content

feat: expose eval, agent context, and validation tools on /mcp#2007

Merged
bokelley merged 3 commits intomainfrom
bokelley/mcp-eval-tools
Apr 8, 2026
Merged

feat: expose eval, agent context, and validation tools on /mcp#2007
bokelley merged 3 commits intomainfrom
bokelley/mcp-eval-tools

Conversation

@bokelley
Copy link
Copy Markdown
Contributor

@bokelley bokelley commented Apr 8, 2026

Summary

  • Promotes 10 internal Addie tools to first-class MCP tools on the /mcp endpoint so external clients (Claude Code, Claude Desktop, custom agents) can call them directly instead of routing through chat_with_addie
  • Evaluation tools (4): probe_adcp_agent, evaluate_agent_quality, test_rfp_response, test_io_execution
  • Agent context tools (3): save_agent, list_saved_agents, remove_saved_agent — completes the eval workflow by allowing credential storage
  • Validation tools (3): validate_json, get_schema, validate_adagents — stateless schema/property validation

Design decisions

  • Member tools (eval + agent context) bridge MCPAuthContextMemberContext per-request, with an auth gate that rejects anonymous callers (isError: true)
  • Stateless tools (schema + property) are created once at startup
  • Tool definitions are pulled from existing arrays (MEMBER_TOOLS, SCHEMA_TOOLS, PROPERTY_TOOLS) — no duplication
  • Startup validation asserts all exposed tool names exist in their source arrays

Closes #2004

Test plan

  • npx tsc --noEmit — clean compile
  • 563 unit tests pass (pre-commit hook)
  • 23 dedicated unit tests for exposed tools (definitions, schemas, auth rejection)
  • Local server test: tools/list returns 32 tools, all 10 new tools present with correct schemas
  • Auth-required tools return isError: true for anonymous callers
  • Code review: addressed must-fix (auth gate, type assertion) and should-fix items
  • Security review: addressed SSRF prevention (M-1) and auth bridging concerns (M-2)

🤖 Generated with Claude Code

bokelley and others added 2 commits April 8, 2026 16:07
Promotes 10 internal Addie tools to first-class MCP tools so external
clients (Claude Code, Claude Desktop) can call them directly instead of
routing through chat_with_addie.

- Evaluation (4): probe_adcp_agent, evaluate_agent_quality,
  test_rfp_response, test_io_execution
- Agent context (3): save_agent, list_saved_agents, remove_saved_agent
- Validation (3): validate_json, get_schema, validate_adagents

Member tools bridge MCPAuthContext → MemberContext per-request and
require authentication (anonymous callers get isError: true).
Stateless tools (schema/property) are created once at startup.

Closes #2004

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Comment thread server/src/mcp/server.ts Fixed
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@bokelley bokelley merged commit 93e272f into main Apr 8, 2026
12 checks passed
@bokelley bokelley deleted the bokelley/mcp-eval-tools branch April 8, 2026 23:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Expose evaluate_agent_quality, test_rfp_response, test_io_execution as direct MCP tools on /mcp endpoint

1 participant