nodecode-mcp-router

MCP Router: pick the MCP tools that fit a query inside a context budget, and describe each choice. A Nodecode rewrite of Pinutss/jev-mcp-router by Pinutss.

With it the model has one more tool:

mcp_select {"query": "read a local file, then search GitHub issues",
            "tools": [{"id": "filesystem_read", "description": "Read local files.",
                       "tags": ["files", "read"], "tokens": 80,
                       "permissions": ["read_files"], "fallback_id": "docs_search"},
                      {"id": "github_search", "description": "Search GitHub issues.",
                       "tokens": 120, "permissions": ["read_github"]}],
            "max_tools": 1, "budget_tokens": 250}
;; => {"decision": "select",
;;     "selected": [{"id": "filesystem_read", "score": 0.866, "tokens": 80,
;;                   "reasons": ["relevance=1.00", "tokens=80", "efficiency=0.56",
;;                               "max_calls=8"], "permissions": ["read_files"], ...}],
;;     "fallback": {"id": "docs_search", ...},
;;     "total_tokens": 80,
;;     "reasons": ["best_match", "relevance=1.00", ...],
;;     "rejected": [{"id": "github_search", "reason": "low_relevance"}]}

and the same selection from eval:

(mcp-router:select-tools "read a local file"
                          (list (mcp-router:make-tool-profile
                                 :id "filesystem_read" :name "Filesystem Read"
                                 :description "Read local files." :tokens 80)))

When many MCP tools are available and only a budgeted subset should be offered, the tool ranks the catalog by the words it shares with the query, fits the best into max_tools and budget_tokens, explains each choice, names every rejection with its cause, abstains when nothing clears the floor, and previews one fallback. It never runs a tool and it never dials: the ranking is lexical and local, so the same catalog and the same query always answer the same selection.

Typed as a command, /mcp-router says what the cell does and, given a query, selects over a small demo catalog so the ranking is visible without a config. (help :mcp-router) carries the manual while the cell runs, and every request's help section carries one line naming the tool.

Get it

Tell the agent "add the mcp-router cell", or run nodecode add nc://mcp-router.

Set it up

Nothing is required: the selector is local and needs no key. The section is on unless it says "enabled": false, and each member has a default:

{
  "mcp-router": {
    "enabled": true,
    "max_tools": 8,          // how many tools a selection may keep
    "budget_tokens": 800,    // the context budget the kept tools must fit
    "min_relevance": 0.15,   // the lexical floor, 0.0 to 1.0
    "omit_rejected": false,  // true to leave the rejected list out of the answer
    "extra_stopwords": []    // words the lexical match ignores
  }
}

max_tools, budget_tokens and min_relevance are the ceiling and the floor a call that names neither gets; a call's own members win. A min_relevance outside 0.0 to 1.0 refuses the start, and the folder's line in /cells says so until it is fixed.

Use it

A catalog of MCP tools. Pass what your MCP servers offer — id, name, description, server, tags, tokens, permissions, enabled, fallback_id, max_calls, scope (all but id optional) — and the query you are about to run.
A budget you must not exceed. max_tools and budget_tokens cut the list; what was cut is in rejected with max_tools or budget beside it.
A decision you can read out. reasons on each kept tool (relevance, tokens, efficiency, max_calls) and on the answer itself; abstain_reason when nothing was kept (no_candidate, no_fallback, fallback_disabled).
A tool that just failed. failed_tool_id makes it a replacement choice: the failed tool's own fallback_id is tried first, and the decision reads fallback.
Policies before scores. required_permissions and forbidden_permissions are declared by the caller: a candidate lacking a required permission, or holding a forbidden one, is rejected whatever its words. Selection is not authorization — a permission is what the catalog declares, and a query, a tool or a model cannot add one.

What it does not do

No execution. It picks and explains; the caller runs the tool.
No remote judge. The original can hand the ranking to JEV or to an LLM gateway vendor when keys are configured (JEV_*, GATEWAY_*); this cell is the local mode, which is the original's default. There is no key, no endpoint and no network call.
No HTTP server and no demo page. The original also serves /v1/select and a demo page; here the tool, /mcp-router and (mcp-router:select-tools ...) from eval take that place.
No Windows-specific work: nothing here touches the filesystem, a process or a port.

Test

.github/test-cell.sh tools/nodecode-mcp-router

MIT licensed; see NOTICE.