trace-mcp

---
# Front matter only so Jekyll renders the {{ site.data.counts.* }} tags below
# (TRA-263). Stripped from the served file.
layout: null
---
# trace-mcp

> trace-mcp is a framework-aware MCP server that serves precomputed code intelligence — a dependency graph, full-text search, impact analysis, and decision memory — to cut token usage for AI coding agents. It understands {{ site.data.counts.frameworks }} frameworks across {{ site.data.counts.languages }} languages and exposes {{ site.data.counts.tools }} MCP tools over stdio or HTTP.

## Docs

- [Code graph MCP server](https://trace-mcp.com/code-graph-mcp.html): What a code graph MCP server is, how persistent symbol, call and framework relationships reduce agent prompt bloat, and how trace-mcp compares to alternatives.
- [Tools reference](https://trace-mcp.com/tools-reference.html): Full list of the {{ site.data.counts.tools }} MCP tools and {{ site.data.counts.resources }} resources trace-mcp exposes, registered dynamically per detected framework.
- [Tool index](https://trace-mcp.com/tools-index.html): Alphabetical index of every MCP tool with its one-line description, generated from the server's own registrations.
- [What `trace init` installs](https://trace-mcp.com/what-trace-init-installs.html): What `trace init` writes to your machine — the guard hook, the opt-in Read/Bash mirrors, the routing block, client config, and the tweakcc tiers — plus how to audit your own startup block.
- [Serena, Repomix and 20+ code graph MCP servers compared](https://trace-mcp.com/comparisons.html): Feature-by-feature comparison against other code-graph and code-intelligence MCP servers.
- [trace-mcp vs Repomix](https://trace-mcp.com/vs/repomix.html): Head-to-head against Repomix — packing a repository into one prompt vs indexing it into a queryable graph, and when Repomix is the better pick.
- [trace-mcp vs Serena](https://trace-mcp.com/vs/serena.html): Head-to-head against Serena — a live LSP proxy vs a precomputed framework-aware graph.
- [trace-mcp vs codebase-memory-mcp](https://trace-mcp.com/vs/codebase-memory-mcp.html): Head-to-head against the closest peer — same knowledge-graph premise, opposite bets on language breadth vs framework depth.
- [trace-mcp vs CodeGraphContext](https://trace-mcp.com/vs/codegraphcontext.html): Head-to-head against the peer that drives the most SCIP indexers — eleven indexer families and a graph database you choose vs one embedded store, framework edges and a write path.
- [trace-mcp vs SocratiCode](https://trace-mcp.com/vs/socraticode.html): Head-to-head against SocratiCode — 25 specialized agent tools, Docker-hosted vector search and Ollama embeddings vs zero Docker overhead, sub-millisecond local graph querying and framework-aware edge extraction.
- [trace-mcp vs jCodeMunch](https://trace-mcp.com/vs/jcodemunch.html): Head-to-head against jCodeMunch — 70+ AST languages and a 3-tool Counter front door under a non-commercial license vs permissive MIT, 81 languages, framework edges, and safe AST refactoring.
- [trace-mcp vs TokenSave](https://trace-mcp.com/vs/tokensave.html): Head-to-head against TokenSave — 86 advertised tools in Rust and string replacement primitives vs permissive MIT, 88 framework integrations, 5-tier call graphs, and AST refactoring.
- [trace-mcp vs Narsil-MCP](https://trace-mcp.com/vs/narsil-mcp.html): Head-to-head against Narsil-MCP — 90 advertised tools, SPARQL/RDF triple stores and 4-tier CCG disclosure vs permissive MIT, 88 framework integrations, 5-tier call graphs, and AST refactoring write tools.
- [trace-mcp vs GitNexus](https://trace-mcp.com/vs/gitnexus.html): Head-to-head against GitNexus — LadybugDB graph database, in-memory Leiden community detection and raw Cypher queries under a non-commercial license vs permissive MIT, 88 framework integrations, typed MCP tools, and AST refactoring.
- [trace-mcp vs Graft](https://trace-mcp.com/vs/graft.html): Head-to-head against Graft — markdown concept cache and external LLM dependency vs local SQLite graph, zero API tokens, 88 frameworks, and AST refactoring.
- [trace-mcp vs codegraph](https://trace-mcp.com/vs/codegraph.html): Head-to-head against the largest peer by stars — one advertised tool optimised for orientation vs a broad surface that continues past navigation.
- [trace-mcp vs Context Mode](https://trace-mcp.com/vs/context-mode.html): Head-to-head against the adjacent lane — compressing what tool calls return vs making structural questions about code cheap to ask.
- [trace-mcp vs code-review-graph](https://trace-mcp.com/vs/code-review-graph.html): Head-to-head against the incremental SQLite graph peer — blast-radius CI analysis and uncertainty handling vs framework awareness and refactoring.
- [trace-mcp vs ast-grep](https://trace-mcp.com/vs/ast-grep.html): Head-to-head against ast-grep — structural pattern search, YAML lint rules and rewrites vs a persistent queryable graph with impact analysis and framework edges.
- [trace-mcp vs Continue](https://trace-mcp.com/vs/continue.html): Head-to-head against Continue — the open-source coding agent for VS Code, JetBrains and terminal vs the persistent code graph any agent queries.
- [trace-mcp vs Aider](https://trace-mcp.com/vs/aider.html): Head-to-head against Aider — terminal pair programming with a repo map and git-native loop vs a persistent graph with stored callers, impact and framework wiring.
- [trace-mcp vs Cody](https://trace-mcp.com/vs/cody.html): Head-to-head against Sourcegraph Cody — enterprise-only cloud search plus assistant vs a local-first graph with no key, no cloud and framework edges.
- [trace-mcp vs ripwire](https://trace-mcp.com/vs/ripwire.html): Head-to-head against ripwire — zero-dependency CLI map with published evals vs a persistent MCP graph with framework edges and AST refactoring.
- [trace-mcp vs IDE context](https://trace-mcp.com/vs/ide-context.html): Head-to-head against Cursor rules and Copilot instructions — authored session guidance vs a derived persistent graph every agent queries.
- [Repomix vs codegraph](https://trace-mcp.com/vs/repomix-vs-codegraph.html): The two other tools compared to each other — packing a repository into one file vs indexing it into a queryable graph, and where trace-mcp sits between them.
- [PR review context benchmark](https://trace-mcp.com/pr-context-benchmark.html): Input-token cost of AI code-review context measured on 60 merged pull requests across six open-source repositories — trace-mcp versus loading the changed files.
- [How to reduce Claude Code token usage](https://trace-mcp.com/reduce-claude-code-token-usage.html): Seven measured tactics for cutting agent token cost, including the ones that have nothing to do with trace-mcp.
- [Configuration](https://trace-mcp.com/configuration.html): The config file options you reach for, explained; trace-mcp works out of the box without one.
- [Config index](https://trace-mcp.com/config-index.html): Every key the config schema accepts, with its type, allowed values and real default, generated from the schema.
- [Architecture](https://trace-mcp.com/architecture.html): The two-pass indexing pipeline and how the dependency graph is built and stored.
- [Supported frameworks & languages](https://trace-mcp.com/supported-frameworks.html): The {{ site.data.counts.languages }} supported languages and the frameworks each plugin understands.
- [Language capability matrix](https://trace-mcp.com/language-matrix.html): What each of the {{ site.data.counts.languages }} supported languages actually extracts — parser, edge types, and test coverage.
- [Session Analytics & Coverage Intelligence](https://trace-mcp.com/analytics.html): The built-in engine that parses agent session logs and tracks token savings and waste.
- [TOON Token Savings — Measured](https://trace-mcp.com/toon-savings.html): Real-world token measurements for the TOON output format across trace-mcp tools.
- [Decision memory](https://trace-mcp.com/decision-memory.html): The persistent decision knowledge graph linking architectural decisions and tradeoffs to code.
- [Daemon memory](https://trace-mcp.com/daemon-memory.html): Measured resident-set attribution for the trace-mcp daemon — what each region holds and which config knob bounds it.
- [Telemetry & Observability](https://trace-mcp.com/telemetry.html): The OpenTelemetry-compatible span emitter for AI provider calls and MCP tool invocations.
- [Privacy](https://trace-mcp.com/privacy.html): Everything the anonymous daily install ping sends, both ways to turn it off, and how to delete the local state trace-mcp keeps.
- [Quality gates](https://trace-mcp.com/quality-gates.html): How `.trace.json` quality-gate thresholds are configured and calibrated.
- [System Prompt Routing via tweakcc](https://trace-mcp.com/tweakcc.html): Wiring trace-mcp tool usage directly into Claude Code's system prompt via tweakcc.
- [Development](https://trace-mcp.com/development.html): Local setup, build, and test instructions for contributing to trace-mcp.
- [Tool response token cost](https://trace-mcp.com/perf/response-tokens/): Per-tool measured response token cost weighted by real call volume, including the tools that cost more than the file reads they replace.

## Optional

- [llms-full.txt](https://trace-mcp.com/llms-full.txt): Full text of every docs page above, concatenated into a single file.