ContextSniper

Make a local codebase searchable from Claude Code or OpenClaw with semantic code search, filtered long outputs, and exact-replacement file edits.

中文文档

Usage Cost Reduction

The SWE Lite verification reports include token and turn breakdowns for Claude Code and OpenClaw runs. We chose text-embedding-3-large as the embedding model.

Claude Code	OpenClaw

Claude Average Per Task

Metric	Legacy	ContextSniper	ContextSniper Filter	ContextSniper Reduction	ContextSniper Filter Reduction
Input tokens	1,534,963	941,761	810,396	-39%	-47%
Output tokens	16,969	11,656	11,946	-31%	-30%
Total tokens	1,551,933	953,417	822,342	-39%	-47%

OpenClaw Average Per Task

Metric	Legacy	ContextSniper	ContextSniper Filter	ContextSniper Reduction	ContextSniper Filter Reduction
Input tokens	1,309,937	769,590	622,321	-41%	-52%
Output tokens	17,439	10,399	12,982	-40%	-26%
Total tokens	1,327,377	779,989	635,303	-41%	-52%

Overview

ContextSniper ships two local plugins:

Host	Plugin	What it adds
Claude Code	claude-plugin/	MCP tools for code search/edit plus prompt policy injection
OpenClaw	openclaw-plugin/	Native tools, prompt policy injection, and read/exec filtering

Both plugins can start the local ContextSniper backend and AGFS service for you, then stop the services they started when the host exits.

Quickstart

Prepare a fresh checkout once:

git clone https://github.com/Calluking/ContextSniper.git
cd ContextSniper
./bootstrap.sh

bootstrap.sh is interactive for setup steps: if Ubuntu packages or Python .venv are missing, it asks before installing anything. It also checks whether OpenClaw and Claude Code are installed, then asks which integration to set up. It does not ask you to type API keys into the installer.

Export your model and embedding settings: (Use deepseek and text-embedding-3-small as example)

export DEEPSEEK_API_KEY="<your-deepseek-key>"
export DEEPSEEK_BASE_URL="https://api.deepseek.com"
export OPENCLAW_MODEL="deepseek/deepseek-v4-flash"
export CONTEXTSNIPER_EMBEDDING_API_KEY="<your-embedding-key>"
export CONTEXTSNIPER_EMBEDDING_BASE_URL="https://api.openai-proxy.org"
export CONTEXTSNIPER_EMBEDDING_MODEL="text-embedding-3-small"
export ANTHROPIC_MODEL="haiku"

Then load the settings:

source setup_env.sh

./bootstrap.sh creates .venv, installs the runtime dependencies in requirements.txt, and builds agfs/build/agfs-server. Full SWE-bench validation is optional; install that heavier stack only when needed with ./bootstrap.sh --install-swe-deps.

Start Claude

From the project you want Claude to edit:

cd /path/to/project
claude --plugin-dir "$CONTEXTSNIPER_DIR/claude-plugin"

Do not pass --mcp-config; the Claude plugin owns its .mcp.json.

Start OpenClaw

From the project you want OpenClaw to edit:

cd /path/to/project
openclaw chat --local

Requirements

• Python 3.11+
• Go 1.22+
• Claude Code CLI, OpenClaw CLI, or both
• An OpenAI-compatible embedding endpoint and API key
• Conda, only if you use the SWE runner's default local validation environment

Verify

Claude:

/plugin

Expected status:

contextsniper Plugin · inline · ✔ enabled
└ contextsniper MCP · ✔ connected

OpenClaw:

openclaw plugins inspect contextsniper --runtime --json

The runtime output should include:

contextsniper_health
contextsniper_index_codebase
contextsniper_search_code
contextsniper_edit_file

To confirm a run used ContextSniper search/filtering:

latest=$(ls -t ~/.openclaw/agents/*/sessions/*.jsonl | grep -v trajectory | head -1)
rg -n "ContextSniper|FILTER IS TRIGGERED|contextsniper_search_code|contextsniper_edit_file" "$latest"

For Claude, inspect the current Claude debug log or the /plugin panel. A healthy code run should show MCP tool names containing search_code and edit_file.

Configuration

Local settings should come from your shell environment or shell profile, which setup_env.sh imports before applying repo defaults. Do not commit real API keys.

Common settings:

Variable	Purpose
CONTEXTSNIPER_EMBEDDING_API_KEY	API key for real semantic code search
CONTEXTSNIPER_EMBEDDING_BASE_URL	OpenAI-compatible embedding endpoint
CONTEXTSNIPER_EMBEDDING_MODEL	Embedding model name
PY_BIN	Optional Python override for plugin/backend runtime
CONTEXTSNIPER_FILTER_ENABLED	Toggle read/exec filtering, default 1
CONTEXTSNIPER_INJECT_FILTERING_PROMPT	Toggle the filtering-strategy prompt section

requirements.txt includes httpx[socks], and setup_env.sh sets NO_PROXY/no_proxy for 127.0.0.1, localhost, and ::1 so local ContextSniper/AGFS traffic bypasses HTTP(S)/SOCKS proxies.

Useful Files

• bootstrap.sh: one-command local setup.
• setup_env.sh: shared environment loader.
• claude-plugin/prompts/code_policy_injection.txt: Claude policy prompt.
• openclaw-plugin/prompts/code_policy_injection.txt: OpenClaw policy prompt.
• docs/assets/readme/: generated README diagrams. Refresh with python3 docs/assets/readme/generate_readme_diagrams.py.
• scripts/SWE/: SWE Lite runners and reports.
• LICENSE: Mulan Permissive Software License v2 (MulanPSL-2.0).

SWE Lite Runner

The SWE runners build a task prompt and start Claude or OpenClaw non-interactively. See scripts/SWE/README.md for the six runner modes, run commands, output locations, and log checks.

License

This project is licensed under the Mulan Permissive Software License v2 (MulanPSL-2.0). See LICENSE.

References

• AGFS: ContextSniper bundles and builds the local AGFS server under agfs/ and uses pyagfs for local memory/file-service operations.