Code intelligence for coding agents: search a repository by meaning, symbols, dependencies, and Git history.
Install Seraph with one command:
python -m pip install seraph-mcpThen tell your coding agent:
Read AGENT_SETUP.md and set up Seraph for this repository.
The agent will register Seraph with its MCP support and check that it connects. Seraph starts in Lite mode, so it is ready without downloading a model.
Switch tiers anytime. Fast and Quality add semantic search with local models.
| Tier | Search | Model download |
|---|---|---|
| Lite (default) | BM25 keyword search | None |
| Fast | BM25 + potion-code-16M embeddings |
62 MB |
| Quality | BM25 + BGE-Code-v1 embeddings | About 6 GB |
Use seraph setup to choose interactively, or install a tier directly:
seraph setup --tier fast --install
seraph setup --tier quality --installRegister a local stdio MCP server with command seraph-mcp. Pass repo as the absolute path to the Git checkout on each tool call. Seraph keeps its index in <repo>/.seraph/; repository contents are not uploaded.
For an isolated uvx install, use command uvx with arguments seraph-mcp. See AGENT_SETUP.md for host-specific setup.
Seraph's MCP tools cover:
- Search:
search_code,search_at_version,search_history,search_worktree. - Understand:
find_symbol,find_dependencies,get_repo_map. - Changes:
compare_versions,get_working_tree_status,prepare_change_context,change_impact. - Index and setup:
index_repository,retrieval_status,set_retrieval_tier.
Seraph parses code into a versioned local index, then combines lexical, semantic, structural, graph, and change-history signals to return relevant code with source context.
On the CoIR AppsRetrieval test split (3,765 queries over 8,765 Python solutions):
| Embedder | nDCG@10 | MRR@10 | Recall@10 | Recall@100 |
|---|---|---|---|---|
| BGE-Code-v1 (submitted) | 0.9770 | 0.9702 | 0.9971 | 0.9992 |
| Qwen3-Embedding-0.6B | 0.7487 | 0.7039 | 0.8882 | 0.9835 |
These are retrieval-quality results on this benchmark, not a measure of agent task success. The benchmark and submission details are in the PRISM 2026 release.
For source checkout and evaluation instructions, see AGENT_SETUP.md and the experiments/ directory.
Team floppydisk · SRM Institute of Science and Technology · Theme 1: Agentic Code Intelligence
MIT. See LICENSE.