Local code search and task context packs for coding agents
io.github.bvolpato/ivygrep — Model Context Protocol (MCP) Server
The io.github.bvolpato/ivygrep MCP server provides local code search and task context packs for coding agents. It is described as a tool for organizing and supplying relevant context derived from code during agent workflows, supporting retrieval of information needed to complete tasks.
🛠️ Key Features
Local code search
Task context packs for coding agents
🚀 Use Cases
Supplying code-derived context to coding agents
Searching locally available code as part of task execution
⚡ Developer Benefits
Server focus on local code search
Context packaging for agent task workflows
⚠️ Limitations
Source description does not specify supported platforms, authentication, or tool interfaces.
No details provided on configuration options or performance characteristics.
Turn coding tasks into bounded, branch-aware context.
One search for discovery. One implementation map from task, current Git changes, and code relationships. Everything stays local.
ivygrep search followed by a bounded task-context pack
# Find code by intent
ig "where is refresh token rotated?"# Build context from code and current changes
ig context "fix refresh-token races" --since main --budget 8000
Search answers where. Context answers what an agent needs to change safely.
The context command combines task anchors with commits since the branch point, staged and dirty files, paths from an issue or trace, and indexed code relationships. It returns one bounded Markdown pack. Every snippet includes its path, lines, role, selection reason, and relationship to the task.
Install
bash
brew install bvolpato/tap/ivygrep
bash
curl -fsSL https://raw.githubusercontent.com/bvolpato/ivygrep/main/install.sh | sh
WinGet submission is awaiting registry approval. crates.io publishing is prepared but not live yet. Use one of the supported installers above until those registries list ivygrep.
Installers select a compatible archive, verify its SHA-256 checksum, install ig, and report the selected backend. Apple Silicon uses Metal. NVIDIA Linux hosts use the Linux x86_64 CUDA build when CUDA 13 and compute capability 8.0 or newer are available. Other systems use portable local inference. Run ig hardware to see detected hardware, compatibility limits, and the matching reinstall command.
Useful controls include -n for result files, -C for context lines, --type for language, --include and --exclude for path globs, plus --lexical-only, --hash, --json, and --no-index. Run ig --help for the full reference.
Search notes and memories
ivygrep also indexes Markdown, text, JSON, and other document files. Precompute local vectors once, then search a notes directory by meaning:
bash
ig --add ~/notes --wait-for-enhancement
ig -n 20 "what did we decide about cache invalidation?" ~/notes
Default daemon-backed queries across CLI, MCP, Web, and TUI blend semantic and lexical retrieval; no semantic opt-in flag is required. For implicit questions whose initial results are overwhelmingly note-like files, ivygrep automatically runs two generic local memory probes concurrently and fuses their ranks. The index stays live as notes change. Queries, note contents, embeddings, and results stay local.
On the public MemoryQuest benchmark, default CLI search retrieved 74.9% of required memories in the top 20 and retrieved every required memory for 44.9% of questions. Warm CLI p95 was 86.05 ms across 535 implicit questions and 3,878 preindexed sessions. Report documents protocol, single-query control, published reference points, and comparability limits.
Connect coding agents
Codex and Claude Code packages install MCP configuration plus focused task-context skill:
bash
codex plugin marketplace add bvolpato/ivygrep
codex plugin add ivygrep@ivygrep
claude plugin marketplace add bvolpato/ivygrep
claude plugin install ivygrep@ivygrep
Automatic setup detects the client, preserves existing configuration, writes the absolute ig path, verifies the MCP handshake, and runs a search:
bash
ig agent install claude
ig agent install codex
ig agent install cursor
ig agent doctor
Restart an open client after installation. Manual MCP setup is also available:
bash
claude mcp add -s user ig -- ig --mcp
codex mcp add ig -- ig --mcp
gemini mcp add --scope user --transport stdio ig ig --mcp
Agents call ig_search for discovery. Set output=context_pack and budget_tokens=8000 when the task needs implementation context. Pass the absolute path to the active repository or worktree. Worktrees reuse the base index and store only changed chunks and tombstones.
Context packs can include definitions, callers, references, dependencies, dependents, tests, configuration, and docs.
Use ivygrep before broad filesystem scans. Pass absolute active worktree path.
Use natural-language queries for concepts and literal=true for identifiers.
For implementation, request output=context_pack with budget_tokens=8000.
Repository context, not another result list
Current work is input.--since main brings committed branch changes, staged files, dirty files, and untracked files into retrieval.
Relationships expand the task. Definitions, callers, references, dependencies, dependents, tests, configuration, docs, and recent co-changes connect likely implementation surfaces.
Token budget is a contract. Context selection, deduplication, and trimming produce one pack sized for the receiving agent.
Evidence stays inspectable. Exact paths and lines, roles, reasons, and relationships explain why every snippet is present.
Worktrees stay thin. Each worktree reuses its repository's base index and stores only divergent chunks and tombstones.
Indexes stay live. Changed chunks update incrementally. Lexical results remain available while neural vectors build in the background.
One context model serves every client. CLI, JSON, MCP, and Web return the same structured evidence.
Search when exploring. Build context when implementing.
ivygrep combines Tantivy BM25, exact lookup, USearch ANN, Tree-sitter chunks, a SQLite relationship graph, local Candle embeddings, and Git-aware incremental indexes. Lexical results are available while neural vectors build in the background. Ranking is deterministic.
ivygrep supports 45 language and file types. Twenty-four use Tree-sitter AST chunking, including Rust, Python, Go, JavaScript, TypeScript, Java, C/C++, C#, Kotlin, Scala, PHP, Ruby, Swift, Elixir, Zig, Bash, Haskell, OCaml, Lua, Dart, Objective-C, Perl, and Starlark.
System performance
On the deterministic one-million-chunk corpus, v1.2.7 median warm CLI p95 is 6.19 ms, controlled indexing reaches 150,576 chunks/s, and the final index is 0.42 GiB across three sequential trials. These are system measurements, not agent outcomes. Hardware, repository shape, model, index state, and load affect results.
Code, queries, embeddings, and indexes stay local. Neural mode downloads pinned model assets once. Use --hash or a hash-only build to avoid model downloads.
ig --web binds to loopback by default. A non-loopback listener prints an authenticated URL but still uses plain HTTP. Use a trusted network, Tailscale, or an encrypted tunnel, and never expose the listener directly to the internet. File contents, including non-ignored dotfiles, can appear in the local index and snippets.
Report vulnerabilities through a private security advisory. Release archives include checksums, SBOMs, and provenance.
