# UCN - Universal Code Navigator

Auditable answers, with every match counted, unsure ones marked and blind spots named.

[![npm](https://img.shields.io/npm/v/ucn)](https://www.npmjs.com/package/ucn)
[![tests](https://img.shields.io/endpoint?url=https://gist.githubusercontent.com/mleoca/0e10a790e16ab61ddd233e05645e203e/raw/ucn-tests.json)](https://github.com/mleoca/ucn/actions/workflows/ci.yml)
[![license](https://img.shields.io/npm/l/ucn)](LICENSE)

If you work with AI Agents, add UCN as a [Skill or MCP tool](#ai-setup). One tool
gives the agent compact, source-linked answers to caller, impact, and test
questions, with uncertainty labeled instead of guessed. The answer depends on
the indexed source, not on who asks: a person at a terminal and an agent over
MCP receive the same records for the same question.

<img src="assets/readme/ucn-connected-scope.gif" width="736" alt="Conceptual overview of UCN's relationship views, multi-hop exploration, visible uncertainty, index reuse, and refresh after edits. Nodes and timing are illustrative, not a captured query or benchmark.">

Conceptual overview · [Still image](assets/readme/ucn-connected-scope.png)

Use the CLI directly, install the [agent skill](#agent-skill-no-server-needed),
or connect through [MCP](#mcp). One engine supplies all three:

```text
  Terminal              AI Agents           Agent Skills
       │                    │                    │
      CLI                  MCP                 Skill
       └────────────────────┼────────────────────┘
                            │
                     ┌──────┴──────┐
                     │ UCN Engine  │
                     │  commands   │
                     │ tree-sitter │
                     └─────────────┘
```

UCN uses tree-sitter abstract syntax trees (ASTs) for static code analysis,
without compiling the project or starting
a language server. The CLI runs on demand and reuses an incremental index;
MCP keeps a process available for repeated queries. No project configuration
is required, and the cache lives outside the repository.

Supports JavaScript, TypeScript, JSX/TSX, Python, Go, Rust, Java, C, C++, C#,
and HTML inline scripts.

## Install

```bash
npm install -g ucn                    # Node.js 20+
```

## In the shell

From a project directory:

```bash
ucn repo
ucn show handleRequest --lines
ucn source handleRequest --raw
ucn impact --staged --lines
ucn check --staged
```

`repo` maps the project. `show --lines` locates callers, `source --raw`
retrieves the implementation, and `impact` and `check` inspect a staged change.
When a name is ambiguous, `find` returns a `file:line:name` handle that
subsequent commands accept.

`--lines` returns `path:line:text` records; `--raw` returns source code.
Both fit an agent's existing scripts without parsing a human-readable report.

Text search remains useful for comments, configuration, strings, and code
outside the supported languages. `usages` provides the literal-name inventory
when the task needs every occurrence, including those that are not calls.

## Code navigation and change analysis

`show` gives a symbol's summary, callers, and callees by default. Select
`--sections` to request source, tests, or related context, or set an output
budget to keep the answer focused. `trace` follows the call graph across files, down into callees
or up through callers toward entry points. Unverified relationships remain
visible, and the tree's accounting reports where exploration stopped.

`impact` connects a symbol or Git diff to its callers. `tests` follows indexed
call and reference paths to identify statically linked tests, including links
several hops away. `plan` previews a rename or signature change with source
locations and review items; it does not edit files. Together these commands
support code exploration, refactoring, and change review from a terminal or
an AI agent.

## What an answer establishes

For caller answers, UCN checks bindings, imports, receiver types, and ownership
to distinguish calls to the selected definition from other uses of its name.
Calls without enough evidence stay visible as unverified, with a reason.
A matching method name alone does not establish which implementation runs;
receiver and ownership evidence determine how the candidate is classified.

In ripgrep at [`82313cf9`](https://github.com/BurntSushi/ripgrep/commit/82313cf95849bfe425109ad9506a52154879b1b1),
the selected `file_name` helper has four confirmed call sites and one unverified
candidate. This animation accounts for all 29 matching lines, alongside import
relationships and same-name definitions drawn from UCN's output.

<a href="assets/readme/ucn-evidence-graph.gif"><img src="assets/readme/ucn-evidence-graph.gif" width="736" alt="Captured ripgrep findings for file_name: 29 matching lines in six files, partitioned into 4 confirmed calls, 1 unverified candidate, 17 non-call lines, 7 other-target lines, and 0 unaccounted. The graph shows selected import relationships and the distinct definitions; motion is illustrative."></a>

Captured findings · [Still image](assets/readme/ucn-evidence-graph.png) · [Data](assets/readme/ucn-evidence-graph.json)

The ACCOUNT line reconciles the observed name occurrences: confirmed calls,
unverified candidates, non-call occurrences, and matches attributed to another
target. Each occurrence is counted in exactly one category, so the totals can
be checked line by line rather than taken on trust. CONTRACT describes the
scope of that accounting. Warnings identify source the index could not cover.
These details survive text truncation for an agent's output budget; a budget
too small to hold them withholds the answer and names the budget that would.

An empty result therefore means something specific about the inspected code.
It cannot establish that reflection, generated code, runtime registration, or
external consumers never reach a symbol. `deadcode` supplies candidates to
investigate; deletion still needs corroboration. A refactor preview still
needs the compiler and tests.

## Accuracy and validation

Release gates compare UCN's answers with independent compilers and language
servers on a ten-repository board of pinned production codebases. The local
November 6, 2026 evaluation, using clangd 18.1.8 for C/C++ and .NET 10.0.403
for Roslyn, recorded these sampled caller results:

| Repository | Pinned commit | Oracle | Symbols sampled | Confirmed precision | In-scope recall |
|---|---|---|---:|---:|---:|
| preact-signals | `e0ce9fdf` | ts-morph | 27 | 100% | 100% |
| httpx | `b5addb64` | Pyright | 50 | 100% | 100% |
| cobra | `ad460ea8` | gopls | 50 | 100% | 100% |
| viper | `528f7416` | gopls | 50 | 100% | 100% |
| ripgrep | `82313cf9` | rust-analyzer | 41 | 100% | 100% |
| clap | `d3e59a9a` | rust-analyzer | 50 | 100% | 100% |
| javapoet | `b9017a95` | JDT LS | 50 | 100% | 100% |
| newtonsoft-json | `4f73e743` | Roslyn | 50 | 100% | 100% |
| cjson | `c859b25d` | clangd | 50 | 100% | 100% |
| fmt | `e424e3f2` | clangd | 50 | 100% | 100% |

That evaluation scored 9,677 confirmed caller edges, all correct;
41 further edges were unscored. It reported zero missing in-scope
oracle edges in both caller and callee answers. The separate value-reference
sample scored 1,241 confirmed sites, all correct, with 4 unscored and zero
missing in-scope sites.
All 8,000 cross-command comparisons agreed.
The default dead-code audit found zero false-dead results among 7 scored
claims; 4 additional claims could not be pinned by the oracle and were unscored.

The samples are deterministic and stratified by reference activity. Confirmed
precision applies to scored claims; recall counts in-scope oracle edges found
in either the confirmed or unverified band. Unverified candidates, oracle
abstentions, and unscored findings remain separate. These measurements do not
establish complete runtime knowledge.

The scheduled board covers 43 pinned repositories: the ten above plus zod,
express, hono, zustand, fastify, dayjs, rich, click, attrs, grpc-go, chi, cursive,
itertools, rayon, rand, gson, jsoup, pino, tanstack-form, tqdm, zap, nom, guice,
dapper, lz4, pugixml, stateless, luxon, arrow, more-itertools, rust-url,
indicatif, and cachetools. A rotating fresh-repository arm checks 13 untuned codebases.
Repositories used to repair a finding move to the pinned board and receive
untuned replacements. A fresh failure fails the weekly run; independent
checks still run and upload their reports.

The [repository manifest](eval/lib/repos.js) records the full commits.
The [Publish workflow](https://github.com/mleoca/ucn/actions/workflows/publish.yml)
gates releases, and the [Eval workflow](https://github.com/mleoca/ucn/actions/workflows/eval.yml)
runs the checks on schedule and on demand. Their run pages provide CI results
and evaluation artifacts. Reproduce the checks locally with the oracle
dependencies installed:

```bash
npm run verify
npm run trust:gate
```

## Commands

| Task | Command |
|---|---|
| Repository orientation and health | `repo [--sections=summary,files,stats,health] [--deep]` |
| Symbol summary and relationships | `show <symbol> [--sections=...]` |
| Definition lookup | `find <name> [--type=type] [--with-source]` |
| Complete literal-name inventory | `usages <name>` |
| Literal, regex, or structural search | `search [term] [--regex] [structural flags]` |
| Exact source extraction | `source <symbol\|file:range>` |
| Call trees: down, up, or to entry points | `trace <symbol> [--direction=...] [--to=entrypoints]` |
| Symbol or Git-diff impact | `impact [symbol] [--staged]` |
| Direct or transitively linked tests | `tests <symbol> [--depth=N]` |
| Signature or pre-commit validation | `check [symbol] [--staged]` |
| Refactor preview | `plan <symbol> --rename-to=...` |
| Imports, importers, and cycles | `deps [file] [--direction=...] [--cycles]` |
| Project or file public API | `api [file]` |
| Runtime and framework roots | `entrypoints` |
| Server/client HTTP surface | `endpoints [--bridge]` |
| Conservative dead-code candidates | `deadcode` |
| Likely missing awaits | `audit-async` |
| Stack-trace frame resolution | `stacktrace <text>` |

`--staged` selects the Git-diff form of `impact` or `check`, without a symbol.

`deps --cycles` groups circular dependencies and, in JS/TS and Python,
distinguishes eager imports from deferred or type-only edges. Enumeration limits are disclosed. `repo`
reports source coverage as well as project structure; its HOT ranking counts
confirmed production callers exactly (a refinement budget only binds on very
large repositories, and is disclosed), and `repo --sections=stats --hot` ranks
by all callers.

`endpoints --bridge` matches server routes and client requests recognized by
its framework extractors. `plan` handles code relationships
such as imports, overrides, and interface or trait methods when ownership is
resolved; ambiguous relationships remain review items.

Run `ucn --help` for flags, or use the
[command reference](.claude/skills/ucn/references/commands.md).

## Shell output

Records and code go to stdout; accounting and notes go to stderr. Unverified
records carry a tab-separated reason. An empty `--lines` listing exits 1, an
error exits 2, and a successful listing exits 0. `--lines` supports `find`, `show`,
`usages`, `search`, and `impact`; `show --lines` lists callers by default.
Use `--json` when the script needs structured fields or a tree result.

Listings are deterministic for a given source tree: a fresh build, a cache
load, and an incremental rebuild return the same records in the same order.
`--lines` listings have no default row cap. Explicit limits disclose what they omit,
and a shell-mode character budget fails before writing partial output.
`source --raw` extracts complete functions and classes unless an explicit
line limit is requested; any resulting truncation is reported on stderr.

Ordinary text-mode command errors also exit 2. JSON keeps exit 0 for successful
empty results and exit 1 for command errors (`meta.ok: false` plus `error`).
An exception inside UCN is reported as `Internal error: ...` (JSON adds
`meta.internalError: true`), never as an answer about the code.
Target-less `check` exits 1 when `TRUST` is `BLOCKED`, 0 for other completed
checks, and 2 if it could not run. `check <symbol>` exits 1 for a signature
mismatch or an unverified call site.

Outside `--lines`, `find`, text `search`, `deadcode`, `api`, and
`repo --sections=files` default to a maximum of 500 results. `find` text shows
the top 5 in detail and counts the rest (JSON returns up to 500); `--limit=N`
raises both. Use `--limit=N` to request more; `usages` lists every site,
test files included, unless a limit or `--exclude-tests` is given. Broad `find` queries select candidates by approximate usage
totals before calculating definition-pinned caller activity, and disclose
that selection when limited.

Files named `*.min.js`, `*.bundle.js`, and `*.map` are reported as skipped
sources and make completeness partial. `--include-bundled` (MCP
`include_bundled=true`) indexes the JavaScript bundles while respecting user
exclusions; it bypasses the shared cache. Source maps remain disclosed but
unindexed.

## AI setup

One tool, 18 commands, compact source-linked answers that keep their trust
metadata even when truncated.

### MCP

```bash
# Claude Code
claude mcp add ucn -- npx -y ucn --mcp

# OpenAI Codex CLI
codex mcp add ucn -- npx -y ucn --mcp

# VS Code Copilot
code --add-mcp '{"name":"ucn","command":"npx","args":["-y","ucn","--mcp"]}'
```

<details>
<summary>Manual MCP configuration</summary>

```json
{
  "mcpServers": {
    "ucn": {
      "command": "npx",
      "args": ["-y", "ucn", "--mcp"]
    }
  }
}
```

VS Code uses `.vscode/mcp.json`:

```json
{
  "servers": {
    "ucn": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "ucn", "--mcp"]
    }
  }
}
```

</details>

### Agent Skill (no server needed)

macOS / Linux:

```bash
# Claude Code
mkdir -p ~/.claude/skills
cp -r "$(npm root -g)/ucn/.claude/skills/ucn" ~/.claude/skills/

# OpenAI Codex CLI
mkdir -p ~/.agents/skills
cp -r "$(npm root -g)/ucn/.claude/skills/ucn" ~/.agents/skills/
```

Windows PowerShell:

```powershell
$npmRoot = npm root -g
New-Item -ItemType Directory -Force "$env:USERPROFILE\.claude\skills"
Copy-Item -Recurse "$npmRoot\ucn\.claude\skills\ucn" "$env:USERPROFILE\.claude\skills\"

New-Item -ItemType Directory -Force "$env:USERPROFILE\.agents\skills"
Copy-Item -Recurse "$npmRoot\ucn\.claude\skills\ucn" "$env:USERPROFILE\.agents\skills\"
```

The skill teaches an agent how to orient, pin symbols, choose the smallest
useful command, interpret the evidence tiers, and recover from incomplete
answers. It's guidance over the same engine, not a second implementation.

## Scope

UCN analyzes indexed source in one project. It does not execute the program
or index installed dependencies such as `node_modules` and `site-packages`.
`repo --sections=health --deep` reports source coverage and known analysis
limits.

C and C++ can use `compile_commands.json` for include paths and header context,
but UCN does not run the preprocessor or reproduce a compiler's build-specific
view. C# source generators and external assemblies are also outside the
index. HTML has regression coverage but no compiler/LSP repository oracle.

The CLI, MCP, and skill share the same resolution rules and evidence. Changing
the transport does not change what the engine knows about the code.

---

MIT