# Throughput Lines of Code Per Day — Blog > Engineering posts from Metaphone CMS (brand=ganvil), merged with local build-content bodies. Index: https://throughputlinesofcodeperday.com/blog ## Welcome to Throughput Lines of Code Per Day URL: https://throughputlinesofcodeperday.com/blog/welcome-to-throughput-lines-of-code-per-day Source: metaphone+local Category: engineering Agent-native version control that query-maps your repository instead of dumping files into context. **Metaphor:** A map instead of a filing cabinet. Agents waste most of their context budget rediscovering your repo. Throughput Lines of Code Per Day keeps a queryable code graph inside the VCS, so agents get structured answers instead of file dumps — measured 10.7x context-token savings on real workloads. The repository itself becomes the knowledge surface your tools already speak: Model Context Protocol, outlines, symbol lookup, and impact queries that stay versioned with every commit. Plug Throughput Lines of Code Per Day into your existing agent via MCP and stop paying context tax on every session. --- ## 10.7x Context-Token Savings: How We Measured It URL: https://throughputlinesofcodeperday.com/blog/10-7x-context-token-savings-how-we-measured Source: metaphone+local Category: engineering Every number on our benchmarks page is copied from measured runs — not marketing fiction. **Metaphor:** A stopwatch on the critical path, not a slide deck. Token savings claims are easy to invent and hard to defend. We instrumented real agent workflows against Throughput Lines of Code Per Day MCP surface: API outline generation, single-function body retrieval, caller mapping, transitive change impact, ranked multi-file search, structural pattern match, context bundle packaging, workspace diff, and attribution traces. The counterfactual is what agents do without a graph — grep plus whole-file reads. On those paths, agents spent roughly an order of magnitude fewer input tokens through outline and query tools than through file dumps. Graph queries are the outliers: callers and diffs return lists of names where the alternative is many whole files. Read the full tables on /benchmarks, then run the same tools against your own workspace. --- ## Why Sidecar Code Graphs Go Stale URL: https://throughputlinesofcodeperday.com/blog/why-sidecar-code-graphs-go-stale Source: metaphone+local Category: engineering Sidecars proved agents need a graph. They also proved a second source of truth is a liability. **Metaphor:** A second map that never quite matches the terrain. Sidecar graph tools proved the demand, then went stale between runs. An indexer that lives outside version control is blind to history, missing mid-session edits, and gone when the folder moves. Throughput Lines of Code Per Day computes the graph inside the VCS — updated sub-second per file on every change, keyed by object ID so it is queryable at any commit, not just the last time an indexer happened to run. Agents propose; history stays yours. Delete the derived layer and nothing canonical is lost. Prefer tools that treat the repository as the source of truth — not a cache that drifts. --- ## Plug MCP into Cursor and Claude in 30 Seconds URL: https://throughputlinesofcodeperday.com/blog/plug-mcp-into-cursor-and-claude-in-30-seconds Source: metaphone+local Category: engineering Zero-friction setup: open Model Context Protocol, one config block, agents start querying structure. **Metaphor:** A power outlet, not a custom adapter farm. Integration tax kills good tooling. Throughput Lines of Code Per Day implements the open Model Context Protocol standard. Connecting it to Cursor, Claude Code, Claude Desktop, or any MCP client is a single server entry — local CLI for offline outline mode, or a hosted index token when you want cloud embeddings and multi-agent coordination. No proprietary SDK lock-in. No second repo format. Agents that already speak MCP get outline, search, get_symbol, and context_pack without reinventing discovery. Open the Zero Friction Setup section on the home page and paste the config for your client. --- ## Byte-Perfect Git Compatibility (Not Mostly Compatible) URL: https://throughputlinesofcodeperday.com/blog/byte-perfect-git-compatibility Source: metaphone+local Category: engineering The bridge round-trips exactly — import then export reproduces the original repository. **Metaphor:** A translator that never paraphrases. Security reviews always ask whether it is actually git-compatible. Throughput Lines of Code Per Day is a 1:1 git drop-in for agent workflows. The import/export path is tested for byte-perfect round trips — not aspirational compatibility. Your history, remotes, and review process stay familiar. What changes is what agents can ask of the repository: structural queries instead of recursive greps, symbol bodies instead of whole-file dumps, impact scope instead of hope. Keep git as the wire format. Add a query surface agents can trust. --- ## Outline Mode Beats File Dumps URL: https://throughputlinesofcodeperday.com/blog/outline-mode-beats-file-dumps Source: metaphone+local Category: engineering Agents should read the map before they read the library. **Metaphor:** A table of contents before the novel. Dumping whole files into context is the default — and the most expensive habit. Outline mode hands agents the shape of the workspace first: modules, symbols, imports, and relationships. Only then do they fetch the bodies they actually need via get_symbol or a tight context_pack. That ordering is why measured sessions drop about 90 percent of context-acquisition tokens. Prefer context_pack for a task and outline before monorepo greps. Prefer package-scoped dependency queries over vendored tree walks. Make outline the first tool call in every agent session. --- ## Agent-Native VCS Architecture URL: https://throughputlinesofcodeperday.com/blog/agent-native-vcs-architecture Source: metaphone+local Category: engineering Version control was built for humans reviewing diffs. Agents need a different primary interface. **Metaphor:** A control plane for code, not just a log of patches. Chat-first coding agents inherited git as a storage layer and a text dump as a context strategy. An agent-native VCS still stores history the way teams expect — commits, branches, protection rules — but exposes a query plane optimized for tools: structural search, symbol resolution, change impact, provenance, and packaging of only the slices an agent needs for the current task. Throughput Lines of Code Per Day sits on that boundary: warm forge for canonical truth, phosphor for derived intelligence. AI proposes; humans merge; the graph rebuilds from truth. Build agents that query the repo like a system of record, not a bag of files. ---