trace-mcp

Why STEADY

STEADY (77) reflects a coherent, genuinely agent-native product with a strong open-source and security posture (MIT, local-first, zero egress, audited writes), documented broad framework/language coverage, a clean no-key install path, and real npm distribution. It is not scored higher because this is a public-surface review (medium confidence) — Hlido did not install trace-mcp or run it against a live codebase — and the headline token-reduction claims are asserted without a published, reproducible benchmark, alongside modest community traction (101 stars) and some internal inconsistency on the surface.

Public-surface checklist

• PASS Homepage loads (required) — trace-mcp.com returned a full product page (page_title 'trace-mcp — Make AI agents understand your codebase, not just read it')
• PASS Primary value prop (required) — 'Make AI agents understand your codebase — not just read it' — framework-aware MCP server serving precomputed code intelligence
• PASS Cta present (required) — 'GET STARTED IN MINUTES' / 'npm install -g trace-mcp' copy button / 'VIEW ON GITHUB'
• PASS Pricing or access — Free and open source (MIT license), no signup and no API keys required
• PASS Evidence or demo — Install walkthrough with example terminal session, cross-language edge trace example, and an optional desktop graph explorer shown; live graph-explorer counters read 0 in capture

What we saw

4 screenshots captured by the Hlido engine during the reviewed run (run-21c397ad7f35ce6c-trace-mcp-com). Our own captures — not vendor marketing material.

×

←→

What it does well

• Agent-native by design: an MCP server exposing 169 tools over stdio/HTTP that return answers, not files for the agent to re-read
• Strong trust and security posture — MIT-licensed, runs fully local, no API keys, bundled embeddings, 'no data leaves your machine', dry-run defaults and a guard hook on writes
• Broad documented coverage: tree-sitter parsing plus optional LSP enrichment across many languages, with framework-aware edges for common stacks (Laravel, Django, Next.js, Rails, NestJS, and more)
• Low-friction adoption — a three-command install with no signup, auto-detection of installed MCP clients, and incremental file-watcher reindexing
• Real distribution signal on the surface (~3.1k npm downloads last month) for a young open-source project

What it fails at

• Headline efficiency claims (~40–50% average, up to 94–99%, up to 2× context) are attributed only to unspecified 'baseline agent runs' with no linked methodology, dataset, or independent reproduction
• Internal inconsistency on the public surface — integration counts given as both 58 and 87, and the live graph-explorer counters (languages/integrations/tools) read 0 in Hlido's capture
• Category-creation marketing ('the execution layer for AI systems', 'recomputation → reuse') runs well ahead of demonstrated traction (101 GitHub stars, 2 watching)
• No hands-on behavioral run by Hlido — token savings, tool reliability, and indexing accuracy are asserted by the project, not independently confirmed here
• The heavier features (LSP enrichment, apply_codemod, decision memory, desktop app) are documented capabilities whose real-world reliability is unverified from the public surface

Pricing & access

• ModelOpen source
• Free entry pointYes — a free tier or open-source edition is documented
• Pricing findable on the public surfacePASS Free and open source (MIT license), no signup and no API keys required (tested 2026-09-27)

Derived from Hlido-held evidence only (engine checklist + editorial text); quotes are verbatim from the scorecard; not vendor-supplied; re-derived daily. Verify current prices on the vendor's pricing page. Last verified 2026-09-27.

Related agents

• Claude Code Coding · 91
• Aider Coding · 90
• @aliasjeff/acli Coding · 90

Agent relevance

API CLI MCP SDK Behavioral-testable

Native MCP server: install the CLI (npm install -g trace-mcp), run 'trace-mcp init' to wire it into an MCP client, then 'trace-mcp add' per repo. Agents then call any of the ~169 tools (search, get_outline, find_usages, get_call_graph, get_change_impact, apply_codemod, …) over stdio or HTTP against a locally indexed graph. An open plugin API lets developers add framework-specific edges. Maximally agent-native — MCP is the primary and intended consumption surface.

Agent-friendly score: 9/10

Evidence

• Framework-aware MCP server that serves a precomputed dependency graph, full-text search, impact analysis and decision memory to cut token usage for AI coding agents — source (2026-09-27) verified
• Open source, MIT-licensed, runs locally with no API keys — 'no data leaves your machine', bundled ONNX embeddings, fully offline after first install — source (2026-09-27) verified
• Installs via 'npm install -g trace-mcp' (Node.js 20+) and exposes 169 MCP tools over stdio or HTTP to Claude Code, Cursor, Windsurf, VS Code and other MCP clients — source (2026-09-27) verified
• ~40–50% fewer tokens on average, up to 94–99% in structured workflows, up to 2× effective context — attributed only to 'baseline agent runs', no published methodology — source (2026-09-27)
• Code-linked decision memory and agent-behavior rules keep reuse the default across sessions — source (2026-09-27)