Skip to content

P6-mcp-interfaz.md

Latest commit

 

History

History
42 lines (31 loc) · 5.25 KB

P6-mcp-interfaz.md

File metadata and controls

42 lines (31 loc) · 5.25 KB
Raw

P6 — MCP como interfaz, no como producto

Verdict: 55% — 8 MCP tools are exposed (3 of them graph-aware, added in phase 2) and 11 CLI commands are registered. The parity table below shows where each MCP tool maps to a CLI subcommand and where the gap is. The remaining 45% is the 4 MCP-only tools (vec_read_lines, vec_find_callers, vec_find_dependents, vec_trace_imports), the cross-workspace aggregation in vec_status (MCP-only), and the routing=auto heuristic that chooses dense vs. graph per query (MCP-only).

Verdict

VectorCode ships an MCP server and a CLI; the two are not full peers. The MCP surface is the rich one: 8 tools (vec_search, vec_status, vec_reindex, vec_read_lines, vec_outline, vec_find_callers, vec_find_dependents, vec_trace_imports) defined in src/mcp/handler.rs. The CLI surface is broader on orchestration (init, install, uninstall, upgrade, benchmark, bench-store) and narrower on read operations. There is no vec_read_lines CLI subcommand, no vec_find_* CLI subcommand, and no vec_trace_imports CLI subcommand — graph-aware and snippet-aware reads are MCP-only. Multi-workspace users get a cross-workspace vec_status only via MCP (src/mcp/handler.rs:367 aggregates AppInnerState across all workspaces); the CLI vectorcode status is per-project. The routing=auto heuristic that picks dense vs. graph per query lives in the MCP path (src/engine/router.rs); CLI users must force --mode graph explicitly.

Evidence

  • 8 MCP tools — src/mcp/handler.rs:179 vec_search, :367 vec_status, :409 vec_reindex, :480 vec_read_lines, :540 vec_outline, :600 vec_find_callers, :643 vec_find_dependents, :690 vec_trace_imports.
  • 11 CLI subcommands — src/cli/mod.rs:50-73 (the Commands enum): init, index, search, outline, status, serve, install, uninstall, upgrade, benchmark, bench-store.
  • MCP↔CLI parity table — see below.
  • Cross-workspace vec_status — src/mcp/handler.rs:367 aggregates across AppInnerState for every entry in state.workspaces. The CLI vectorcode status resolves one project root via src/cli/mod.rs:103-129 and never sees the multi-workspace map.
  • routing=auto heuristic — src/engine/router.rs exists and is used by the MCP vec_search path. CLI users can --mode graph | dense | sparse | hybrid | hybrid-rerank (per src/cli/mod.rs:111-114 in the README usage block) but cannot delegate the choice to the engine.

Parity table

MCP tool CLI equivalent Notes
vec_search vectorcode search Full parity (CLI supports --mode, --limit, --threshold, …).
vec_status vectorcode status CLI is per-project; MCP aggregates across workspaces.
vec_reindex vectorcode index --full Flag, not a subcommand. CLI also has vectorcode index (incremental).
vec_outline vectorcode outline <file> Full parity.
vec_read_lines none MCP-only. CLI has no read_lines subcommand.
vec_find_callers none MCP-only. CLI has no find_callers subcommand.
vec_find_dependents none MCP-only.
vec_trace_imports none MCP-only.

Known limits

  • 4 MCP tools have no CLI equivalent — vec_read_lines, vec_find_callers, vec_find_dependents, vec_trace_imports. A user who wants to script a graph traversal or read a 50-line snippet from the shell cannot do it without an MCP client.
  • Cross-workspace status is MCP-only — multi-repo users (the fase-3 deliverable) get the cross-workspace view only via MCP. The CLI vectorcode status is single-root.
  • routing=auto is MCP-only — the engine knows when a question is structural ("who calls X?") vs. conceptual ("where is auth handled?"), but the CLI forces the user to pick. The CLI does not have a --mode auto flag.
  • MCP vec_status schema is JSON-string-typed — it returns a String (JSON-encoded) rather than a structured response. Agents parse it manually. The 8 tools have the same shape; this is a constraint of the rmcp macro shape.
  • No streaming in MCP — long searches block until done. A streaming vec_search is a non-trivial addition.

Links

  • docs/benchmarks.md — the bench harness exercises the same code path the CLI uses (cargo run -- benchmark), independent of MCP.
  • ADR-0001 — the Store port makes both the CLI and the MCP server share the same hot path.
  • Related: P3 (the 3 graph tools come from P3's graph port) · P5 (the SearchStrategy port makes the CLI↔MCP parity trivial in principle).