Open source experimental data plane for ContextForge - an MCP gateway/router that forwards tool calls to backend MCP servers with plugin support, authentication, and observability
Server exposes 6 tools with significant definition gaps. Tool names follow verb_noun convention (call_tool, get_token, configure_user), which is positive. However, parameter descriptions are minimal or absent for most tools. Input schemas are present but lack depth, many parameters have no type constraints, enums, or validation guidance. Output schemas are not documented. Error handling is not evident from the tool definitions. Security concerns: get_token and get_custom_token expose JWT generation as tools (should be server-side only); configure_dataplane is testing-only but still registered. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (WRITE vs READ_ONLY).
Routes and executes tool calls to backend MCP servers with pre/post-hook plugin support, request validation, and progress tracking
Configures global data plane settings by storing GlobalConfig in Redis (testing only, requires with_tools feature)
Configures user-specific settings by storing UserConfig in Redis for a given user ID
Generates a JWT token with custom claims provided in request body, adding standard claims (exp, nbf, iat, jti)
Returns the JWKS (JSON Web Key Set) for token verification with public key in RS256 format
Generates a JWT token for a user with specified tenant and user IDs, optional email, and standard claims (exp, nbf, iat, jti, scopes, permissions)
Token generation tools (get_token, get_custom_token) expose JWT creation as agent-callable tools. This is a security anti-pattern, tokens should be generated server-side and injected into requests, never exposed as tool parameters. Agents log all tool calls; tokens in responses enter LLM context and risk leakage.
Parameter descriptions are minimal or missing. 'claims' in get_custom_token is described as 'Custom JWT claims as JSON object', no guidance on required fields, format, or constraints. 'user_config' in configure_user lacks any description of expected structure. LLMs cannot infer valid input without explicit parameter documentation.
No output schemas documented for any tool. Agents do not know what fields to expect from responses (e.g., does get_token return {token, expires_at, type}? Does get_jwks return {keys, algorithm}?). Without documented output structure, agents cannot reliably chain tools or extract required data.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. call_tool and configure_* are marked WRITE; get_* are marked READ_ONLY. These hints should be present in tool definitions to guide agent planning and safety checks.
configure_dataplane is documented as 'testing only, requires with_tools feature' but is still registered as a callable tool. Testing-only tools should not be exposed in production MCP servers. Either gate it behind a feature flag at runtime or remove it from the tool list.