Get started with TLCPD MCP

TLCPD MCP is a Model Context Protocol server that gives your AI coding agent structured access to your codebase — outlines, exact symbol bodies, call graphs, and ranked context — instead of dumping whole files into the context window.

The result: agents find the right code in fewer, sharper steps and spend ~75% fewer input tokens getting there — roughly 4x more room for actual reasoning.

Two ways to run it

- Hosted — point your agent at your workspace's MCP URL with an access token. Nothing to install; your repo is indexed in the cloud. Best for getting started.

Local — run the tlocpd CLI's MCP server on your machine against a checkout. Your code never leaves your laptop.

> Note: Both surfaces speak the identical MCP tool contract, so you can start hosted and move local (or run both) without changing how your agent calls tools.

The fastest path to your first tool call is the hosted MCP — no install required.

1. Get your MCP URL and token

1. Sign in to your TLCPD dashboard.

Track a repository so it gets indexed.

Open Settings > MCP and copy your hosted MCP URL and a developer token (prefixed gv_pat_).

2. Connect your agent

Add TLCPD MCP using the URL and token above. For Claude Code:

claude mcp add --transport http tlcpd <your-mcp-url> \
  --header "Authorization: Bearer <your-token>"

3. Ask a question that needs the codebase

With the server connected, ask something like *"outline the auth module"* or *"who calls charge_card?"* Your agent will call TLCPD MCP's tools instead of grepping and reading whole files.

> Note: For Cursor, VS Code, Continue, and local setup, see Connect your coding agent below.

TLCPD MCP works with any MCP-capable agent. Grab your hosted MCP URL and gv_pat_ token from Settings > MCP, then use the snippet for your tool.

Claude Code

Remote (hosted):

claude mcp add --transport http tlcpd <your-mcp-url> --header "Authorization: Bearer <your-token>"

Or commit a project-shared .mcp.json:

{
  "mcpServers": {
    "tlcpd": {
      "type": "http",
      "url": "<your-mcp-url>",
      "headers": { "Authorization": "Bearer <your-token>" }
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project; the project file wins):

{
  "mcpServers": {
    "tlcpd": {
      "url": "<your-mcp-url>",
      "headers": { "Authorization": "Bearer <your-token>" }
    }
  }
}

VS Code

VS Code uses a top-level servers key in .vscode/mcp.json:

{
  "servers": {
    "tlcpd": {
      "type": "http",
      "url": "<your-mcp-url>",
      "headers": { "Authorization": "Bearer <your-token>" }
    }
  }
}

Continue

Add a YAML file at .continue/mcpServers/tlcpd.yaml:

name: tlcpd
type: streamable-http
url: <your-mcp-url>
requestOptions:
  headers:
    Authorization: Bearer <your-token>

Local (any agent)

Install the CLI, then run the MCP server against a checkout.

macOS / Linux:

curl -fsSL https://throughputlinesofcodeperday.com/install.sh | sh

Windows (PowerShell):

irm https://throughputlinesofcodeperday.com/install.ps1 | iex

Then import a repo and start the server:

tlocpd import .
tlocpd mcp

Then point your agent's launch config at it:

{
  "mcpServers": {
    "tlcpd": { "command": "tlocpd", "args": ["mcp"] }
  }
}

> Note: The unified tlocpd CLI is rolling out; on older installs the local command is gan mcp. Agent config formats change over time — if a snippet drifts, check your agent's MCP docs for the current shape.

Every tool returns high-signal structure, not raw file dumps — that is where the token savings come from.

Navigation

- outline — the shape of a file or directory (symbols, signatures, nesting) without bodies. Call this before reading a whole file.

get_symbol — the exact source of one symbol (function, type, method). Pull bodies only for what you will edit or quote.

search — semantic + lexical search across the repo; returns ranked locations, not whole files.

structural_grep — pattern search over syntax (AST-aware), not plain text.

Understanding change

- callers — who calls a symbol (reverse call graph).

impact — what a change to a symbol could affect, transitively.

symbol_diff — how a symbol changed between two revisions.

blame_symbol — the history behind a specific symbol.

log_search — search commit history by content and intent.

Task-shaped

- context_pack — give it a task in plain language; it returns a ranked, budgeted bundle of exactly the code the task needs. Start here for open-ended work.

> Note: Prefer context_pack, search, and outline before recursive grep or whole-tree reads — that ordering is what keeps context small.

Both surfaces expose the identical tools; they differ only in where indexing runs.

Hosted

- Your repo is indexed in the cloud; you connect over HTTP with a gv_pat_ token.

Nothing to install; the index stays warm across sessions.

Best for teams and for trying TLCPD MCP quickly.

Local

- The tlocpd CLI indexes and serves a checkout on your machine over stdio.

Your source never leaves your device.

Best for private code and offline work.

> Note: The index is a property of the repository, versioned with it — queries are consistent at any commit, and git round-trips stay byte-for-byte identical.

Hosted access is authorized with a developer token minted in your dashboard.

Get a token

1. Open Settings > MCP in your TLCPD dashboard.

Generate a developer token — it is prefixed gv_pat_.

Pass it to your agent as a bearer header: Authorization: Bearer gv_pat_... (or via the GANVIL_INDEX_TOKEN environment variable for CLI-style configs).

Scope and safety

- Tokens are scoped to your workspace and the repositories it tracks.

Treat a token like a password — anyone holding it can query your indexed code.

Rotate or revoke a token any time from the same Settings page.

> Note: Local (tlocpd mcp) needs no token — it serves a checkout you already have on disk.