This server provides local, bounded MCP tools for indexing a repository and answering code-search, snippet-read, and test-failure questions without dumping whole files into context.
index_repo(root) – Scans a repo, chunks text files into ~80-line windows, and writes a lightweight JSON index at <root>/.context-index/chunks.json; respects .csignore and built-in ignore rules, skips binaries and large files.
search_code(root, query, topK?, maxChars?) – Runs a BM25-style keyword search over the index and returns compact evidence packets (FILE / LINES / SCORE + matching lines with context), capped by maxChars; path-only matches are returned as pointers, and omission markers spell out exact read_snippet calls.
read_snippet(root, path, startLine, endLine) – Reads an explicit line range from one file inside root, capped at 300 lines, with path-traversal protection.
run_test_filtered(root, command) – Runs one of a fixed allowlist (npm_test, pnpm_test, pytest) via spawn with shell: false, and returns only failure-relevant filtered output plus the command and exit code.
CLI subcommands – The same binary can be used as a one-shot shell command (index, search, read, test) instead of an MCP stdio server.
No arbitrary shell execution – Only the four tools are exposed; there is intentionally no run_shell.
Use-case guidance – Use Grep for known identifiers/error strings, and search_code for broad queries that would flood context; follow omission markers with read_snippet to expand only what you need.
Allows running [npm](/mcp/servers/integrations/npm) test via the server, capturing filtered output with failure-relevant lines and exit code.
Allows running [pnpm](/mcp/servers/integrations/pnpm) test via the server, capturing filtered output with failure-relevant lines and exit code.
Allows running [pytest](/mcp/servers/integrations/pytest) -q via the server, capturing filtered output with failure-relevant lines and exit code.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@context-sniper-mcp search code for 'validateUser' in current repo"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
English | 简体中文
A tiny local MCP server that indexes a repository into line-range chunks and
serves compact "evidence packets" (file + lines + score + snippet) instead of
dumping whole files into the model's context. Meant to be shared by
Claude Code and Codex when exploring or debugging a repo. It is a keyword
search with a hard output cap, not a guaranteed token saver: against Grep
it often costs more, see "Token efficiency" below.
No database — the index is a single JSON file written to
<repo>/.context-index/chunks.json.
index_repo { root } — scans root, chunks every text file into
~80-line sliding windows (max 120 lines/chunk), and writes
root/.context-index/chunks.json. There is no extension allowlist: .scss,
.html, .vue, .sh, Makefile and so on are all searchable. Files are
skipped by a built-in gitignore-style rule set (node_modules, .git,
dist, build, .next, coverage, .venv, target, __pycache__,
lockfiles, *.min.*, source maps, .env* secrets, media and archives), by
root/.csignore (see below), by a NUL-byte binary check, and by a
512 KB size cap.
search_code { root, query, topK?, maxChars? } — loads the chunk index
and scores it against query with a BM25-style ranker. topK (default 5,
max 50) bounds the candidate chunks; maxChars (default 6000) bounds the
whole reply. Hits are trimmed to the lines that matched, merged when they
overlap, and returned as FILE / LINES / SCORE + code block; see
"Evidence packets and the output budget" below for the exact shape. If no
index exists yet, it tells you to run index_repo first.
read_snippet { root, path, startLine, endLine } — reads an explicit
line range from one file inside root. Capped at 300 lines per call (longer
ranges are truncated with a note). path is resolved and checked against
root; anything that would escape root is refused.
run_test_filtered { root, command } — runs one of a fixed allowlist of
commands (npm_test → npm test, pnpm_test → pnpm test, pytest →
pytest -q) via spawn with shell: false — no arbitrary shell execution.
Captures stdout/stderr, keeps only lines matching
error|failed|failure|assert|expected|received|traceback or a test-file
path, tail-capped at 120 lines. If nothing matches, falls back to the last
80 raw output lines. Always reports the resolved command and exit code.
There is intentionally no run_shell or equivalent — only the four tools
above are exposed.
Related MCP server: Repo Interrogator
.csignore)Put a .csignore file in the repo root to keep files out of the index. It uses
gitignore syntax: blank lines and # comments are skipped; * matches within
one path segment, ? one character, ** any number of segments, [abc] a
character class; a pattern without / matches a file or directory name at any
depth; a pattern containing / is anchored to the root (a leading / is
optional); a trailing / matches directories only; ! re-includes something
an earlier rule or a built-in default excluded, and the last matching rule
wins. Not supported: \ escapes and nested ignore files.
# .csignore
REVIEW.md
.claude/
docs/**/*.snap
!dist/.gitignore is deliberately not read: version-control noise and search noise
differ (a .gitignore that hides a nested backend repo should not hide it from
search). Re-run index_repo after editing .csignore; the result reports how
many rules were read and how many files or directories they skipped.
search_code never returns whole chunks. For each of the topK candidate
chunks it keeps only the lines containing a query token, plus 2 lines of
context on each side; candidates from the same file whose trimmed ranges
overlap or touch are merged into one hit. So fewer than topK hits can come
back, and no line range is returned twice. A hit looks like this:
[1] FILE: src/output-gate.ts
LINES: 48-74
SCORE: 13.216
```
cwd: root,
shell: false,
env: process.env,
// Run in its own process group so a timeout can reap the whole tree
// (e.g. npm -> node -> test worker), not just the direct child.
detached: true,
... (lines 54-63 omitted)
const timer = setTimeout(() => {
timedOutFlag = true;
// Negative pid targets the whole process group (created via detached).
```LINES is the span of the trimmed hit. Every stretch inside that span that is
not shown is marked in the block:
... (lines A-B omitted) — lines between two matching regions.
... (truncated at 4000 chars; use read_snippet <path> <line> <end> to expand)
— one hit exceeded the per-hit cap; reading resumes at <line>.
... (budget: N chars omitted; use read_snippet <path> <start> <end> to expand)
— the packet's maxChars budget ran out inside this hit (cut at a line
boundary).
A chunk that matched only through its path (say the query payments against
src/payments.ts, whose text never contains that word) is returned as a
one-line pointer instead of code, and only when the file has no text hit in
the packet:
[3] FILE: src/payments.ts
LINES: 1-60
SCORE: 2.1
```
(matched on path only: no query token in lines 1-60; use read_snippet src/payments.ts 1 60 to view)
```The whole packet is capped at maxChars (default 6000, about 1.5k tokens).
Hits are added in score order; the first one that does not fit is cut as
described above, and the remaining hits are listed at the end so they can
still be fetched:
... (budget: 3 more hits omitted: src/cli.ts 81-172, README.zh-CN.md 84-88, README.md 51-62; use read_snippet <path> <start> <end> to expand, or raise maxChars)Whenever at least one hit is shown, that trailer names at least one omitted
hit. Raise maxChars, or narrow the query, to see more of them. A reply of
No matching chunks found. means no chunk contained any query token: the
tokenizer lowercases, splits on anything other than letters, digits and _,
drops single characters, and does not split camelCase, so handleSubmit is
one token.
For installation and client configuration, see INSTALL.md.
The same binary also works as a plain shell command — pass a subcommand and it runs once and exits, instead of starting the MCP stdio server:
context-sniper-mcp index <root>
context-sniper-mcp search <root> <query...> [--top-k N] [--max-chars N]
context-sniper-mcp read <root> <path> <startLine> <endLine>
context-sniper-mcp test <root> <npm_test|pnpm_test|pytest> [--timeout ms]
context-sniper-mcp help
context-sniper-mcp --versionEach subcommand maps 1:1 to the tool of the same purpose above and prints the
same human-readable output. test exits with the underlying test command's
own exit code (or 124 on timeout), so it's usable in scripts, e.g.
context-sniper-mcp test . npm_test || echo "tests failed". Running the
binary with no arguments still starts the MCP stdio server.
Call index_repo once per repo, and again after large changes or an
edit to .csignore.
When you know an identifier, an error string or the file, use Grep (or
Read a short file) as usual. In the measurements under "Token
efficiency" a targeted Grep was 2.5× to 10× cheaper than search_code.
Use search_code when the query is broad enough that Grep would match
across many files and flood the context: its reply never exceeds
maxChars. It is keyword search, so query with words as they appear in the
code, not with a natural-language question.
If a hit ends with an omission marker, follow it: the marker spells out
the exact read_snippet call (<path> <start> <end>, still capped at
300 lines per call) rather than reading the entire file. Raise maxChars
only when the trailer lists several omitted hits you actually need.
If a test fails, use run_test_filtered to get the trimmed failure output instead of piping raw test-runner logs into context.
Paste this into the project's CLAUDE.md or AGENTS.md so the agent knows
when Context Sniper helps and when Grep is cheaper. Every line here is loaded on
every turn, so it is kept short; the tool descriptions the server ships carry
the rest.
## Context Sniper
This repo is served by the `context-sniper` MCP server. Its `search_code` is a
capped keyword search: one reply never exceeds 6000 chars (about 1.5k tokens)
and holds only matching lines plus 2 lines of context. It is not cheaper than
`Grep` for a known identifier; use it for broad queries.
Tools (`root` is always this repo's absolute path):
- `index_repo(root)` — run it if `.context-index/` is missing, and again after
`git pull`, large edits, or a change to `.csignore`. If a snippet's line
numbers no longer match the file, the index is stale: re-run it.
- `search_code(root, query, topK?, maxChars?)` — keyword search (BM25), not
semantic. Query with identifiers, error strings and distinctive words as they
appear in the code; camelCase is one token and single characters are dropped.
Start with the defaults. Raise `topK` for broad queries; raise `maxChars` only
when the trailer lists omitted hits you actually need.
- `read_snippet(root, path, startLine, endLine)` — bounded read, 300 lines max.
Every omission marker in a search result spells out the exact call: copy it
instead of reading the whole file.
- `run_test_filtered(root, command)` — `npm_test` / `pnpm_test` / `pytest`;
returns only the failure-relevant lines.
Workflow:
1. Known identifier, error string or file: use `Grep` / `Read` as usual.
2. Broad query that `Grep` would match across many files: use `search_code`,
then expand with `read_snippet` by following the markers.
3. After editing, verify with `run_test_filtered` instead of the raw test runner.Measured on 2026-09-19; tokens are estimated as characters ÷ 4. Each row is one question, answered four ways:
search_code — the current default (topK 5, maxChars 6000).
Grep -n -C 2 — Claude Code's Grep in content mode, with line numbers
and 2 lines of context, on the one pattern an agent would plausibly try
first (shown in the cell).
Grep files + Read — Grep listing matching files, then Read of
the answering file whole.
Read only — the answering file opened whole, assuming you already
know which file it is.
Read output is counted in its cat -n format (6-wide line number, tab,
line). Corpora: this repo (21 files) and a React + FastAPI project (83 files,
nearly all under 100 lines).
Question | Answering file |
|
|
|
|
|
| 3,186 | 1,236 ( | 5,081 | 5,023 |
|
| 5,996, answer not in the reply | 7,458 ( | 13,699 | 13,536 |
|
| 915 | 1,070 | 593 | 538 |
|
| 2,914 | 297 ( | 1,021 | 974 |
|
| 709 | 272 ( | 634 | 594 |
| none: the word is all over this repo | 5,980 | 58,575 (279 lines in 18 files) | — | — |
All figures are characters. What this shows:
search_code was never the cheapest way to an answer in rows 1–5. Its best row is
__table_name__, where it undercuts Grep -C 2 by 155 chars but costs more
than Grep files + Read of the 13-line file. A targeted Grep beat it
three times, by 2.5× to 10×.
It beats reading a large file whole (rows 1–2), but so does Grep with
context, which is what an agent does without this server.
It is keyword search, not semantic: on the natural-language question in
row 2 it returned src/ignore.ts, README.md, src/search.ts and
test/output-gate.test.mjs, filled the budget, and never reached the answer. Grep only found it
because the pattern guessed the word escape.
A broad word is where the cap pays off: index matches 279 lines across
code, tests, docs and lockfiles, and Grep with context returns ten times
what the capped search_code reply does.
Extra hits cost tokens: all three project queries also return REVIEW.md.
If you did not need that file, those chars are noise.
Two costs the table leaves out. First, the server's four tool definitions
are 3,547 chars of tools/list JSON, and the CLAUDE.md snippet above is
another 1,585; together about 1.3k tokens for every session that loads them,
whether or not a search runs (clients that defer MCP tool schemas until first
use pay only the snippet up front). Second, Claude Code's Edit requires a
Read of the file first, so on any task that ends in an edit, the search
reply is paid on top of that read, not instead of it.
What the server does give you is a bound: one reply never exceeds maxChars,
however broad the query, whereas a loose Grep pattern or a Read of a
large file has no such cap. Whatever was cut can be fetched with the
read_snippet call named in the marker. Use Grep when you know an
identifier or error string; reach for search_code when a query is broad
enough that an uncapped Grep would flood the context.
Atomic index writes — index_repo writes to a temp file next to the
index and rename()s it into place, so a concurrent search_code never sees
a half-written file. The index carries a format version (currently 2); a
file with any other version is treated as missing, and search_code asks you
to re-run index_repo.
Per-process cache — a loaded index is cached by path and mtime for the life of the server process, so repeated searches don't re-parse the JSON; the cache drops the entry when the file changes.
No arbitrary execution — read_snippet resolves path against root
and refuses anything that escapes it; the test runner spawns with
shell: false and a fixed allowlist.
context-sniper-mcp/
├── package.json
├── tsconfig.json
├── src/
│ ├── index.ts # MCP server wiring + tool registration; dispatches to cli.ts
│ ├── cli.ts # shell subcommands (index/search/read/test) for direct CLI use
│ ├── repo-index.ts # scanning, chunking, safe path resolution, index I/O
│ ├── ignore.ts # gitignore-style rules: built-in defaults + .csignore
│ ├── search.ts # BM25-style scoring + evidence packet formatting
│ ├── snippets.ts # bounded, path-safe line-range reads
│ ├── output-gate.ts # allowlisted test runner + output filtering
│ └── tokenize.ts # shared tokenizer used by indexing + search
├── test/ # *.test.mjs unit tests for each src module
├── build/ # compiled output (npm run build)
├── INSTALL.md
├── INSTALL.zh-CN.md
├── README.md
└── README.zh-CN.mdMIT
Scan a repository and build a lightweight chunk index at /.context-index/chunks.json. Run this once before search_code (and re-run after significant code changes).
| Name | Required | Description | Default |
|---|---|---|---|
| root | Yes | Absolute path to the repository root |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations provided, but the description transparently states the tool creates a file at a known path and is read-only in nature (scanning). It lacks details on permissions or safety, but the core behavior is clear.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences: first states action and output, second gives usage context. No unnecessary words, front-loaded with the key purpose.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple index-building tool with one required param and no output schema, the description fully covers what it does, where output goes, and when to run it.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% with a single parameter 'root' described as 'Absolute path to the repository root.' The description adds value by showing the output path relative to root, exceeding schema info.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states 'Scan a repository and build a lightweight chunk index' with a specific output location. It distinguishes itself from siblings like search_code by being a prerequisite indexing step.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicitly advises 'Run this once before search_code (and re-run after significant code changes),' providing clear when-to-use and maintenance guidance.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Read an explicit line range from a single file inside root. Use this only after search_code narrows down a location and the returned snippet isn't enough context. Capped at 300 lines per call.
| Name | Required | Description | Default |
|---|---|---|---|
| path | Yes | File path, relative to root (or absolute, but must resolve inside root) | |
| root | Yes | Absolute path to the repository root | |
| endLine | Yes | 1-based end line, inclusive | |
| startLine | Yes | 1-based start line, inclusive |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Discloses the 300-line cap and that it reads from a file inside root. No annotations, so description carries full burden; could mention read-only nature or return format but fairly transparent.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences, no fluff, front-loaded with purpose then usage. Every word earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple tool with full schema coverage and no output schema, the description adequately covers purpose, usage, and limitations. Could mention return type but not critical.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% so baseline 3. Description adds the line cap but repeats 'line range' which is already in schema. Minimal added value beyond schema fields.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Clearly states it reads a specific line range from a single file inside root, and distinguishes from siblings by referencing search_code for narrowing down locations.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicitly says to use only after search_code narrows a location and when more context is needed, providing a clear condition and alternative.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Run one of a fixed allowlist of test commands (no arbitrary shell) and return only the filtered failure-relevant output: command, exit code, and lines matching error/failed/assert/expected/traceback/test-file-path, tail-capped at 120 lines (falls back to last 80 raw lines if nothing matches).
| Name | Required | Description | Default |
|---|---|---|---|
| root | Yes | Absolute path to the repository root | |
| command | Yes | One of: npm_test, pnpm_test, pytest |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
The description discloses key behaviors: no arbitrary shell, filtered output based on specific patterns, tail-capped at 120 lines with fallback to 80 raw lines. No annotations are provided, so the description carries the full burden and does so thoroughly.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single sentence that is comprehensive but somewhat long. It front-loads the core purpose and includes all necessary details, though it could be slightly more concise.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's simplicity (2 parameters, no output schema), the description fully covers input constraints and output behavior, including matching patterns and fallback logic, making it complete.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% with both parameters described. The description does not add new semantic information beyond what the schema provides, so a baseline of 3 is appropriate.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool runs test commands from a fixed allowlist, not arbitrary shell commands, and returns filtered failure-relevant output. It distinguishes from sibling tools (index_repo, search_code, read_snippet) which have different purposes.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage for running tests and getting filtered output. It does not explicitly mention when not to use or suggest alternatives, but the context is clear enough given the tool's specific purpose.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Search the previously built chunk index with a BM25-style keyword score and return a compact evidence packet (FILE / LINES / SCORE + snippet) instead of full files. Call index_repo first if no index exists.
| Name | Required | Description | Default |
|---|---|---|---|
| root | Yes | Absolute path to the repository root | |
| topK | No | Number of results to return (default 5) | |
| query | Yes | Natural language or keyword query |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Without annotations, the description carries full behavioral disclosure burden. It reveals the search algorithm (BM25-style), output structure (compact evidence packet), and dependency on a pre-built index. It does not discuss failure modes or idempotency, but for a code search tool this level of detail is adequate.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences, front-loaded with the core action and output description. Every part is essential, with zero filler or redundancy.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given no output schema, the description explains the return format (compact evidence packet). It addresses the key prerequisite (index_repo). Missing details about side effects or read-only nature, but overall sufficient for a search tool with three parameters.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema covers all three parameters (100% coverage), baseline is 3. The description adds minor value by mentioning 'BM25-style keyword score' which hints at how the query parameter is used, but it does not elaborate on 'root' or 'topK' beyond schema definitions.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's action: searching a code chunk index using BM25-style keyword scoring. It specifies the return format (compact evidence packet with FILE/LINES/SCORE + snippet) and distinguishes itself from siblings like index_repo (precondition) and read_snippet (full file retrieval).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description explicitly advises to call index_repo first if no index exists, providing a clear prerequisite. Although it does not explicitly state when not to use the tool or mention alternatives, the context from sibling tools and the description's focus on searching versus reading or testing implies appropriate usage.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0index_reporead_snippetrun_test_filteredsearch_codeScored across 4 tools
Each tool has a distinct, non-overlapping purpose: indexing, searching, reading snippets, and running tests. No ambiguity in choosing between them.
All tool names follow a consistent verb_noun (or verb_noun_adjective) pattern in snake_case, making the action and target immediately clear.
Four tools cover the core workflow (index, search, read, test) without extraneous clutter. The count fits the domain perfectly.
The tool set provides a complete loop for code understanding and test failure analysis: build index, search, read context, run tests with filtered output. No obvious gaps for the stated purpose.
Repository knowledge graph MCP server for codebase understanding and debugging.
An MCP server that gives your AI access to the source code and docs of all public github repos
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
