Complementary tool: Pommel for semantic code search

[dbinky]

Hey Steve and Beads community,

I've been using Beads for task tracking and wanted to share a tool I built that addresses a related problem: code discovery.
Beads solves the "what should I work on" memory problem. Pommel tries to solve the "where is the relevant code" problem - so agents can find what they need without loading your entire codebase into context.
How they work together:

Agent checks Beads → "Work on bd-a1b2: Add rate limiting to API endpoints"
Agent searches Pommel → pm search "API request handling" --json --limit 5
Agent gets back specific file:line references instead of grep-ing through 50 files
Context window stays intact for actual coding work

It's a local semantic search tool - vector DB backed by sqlite-vec, embeddings via Ollama, file watcher keeps the index fresh. Nothing leaves your machine.

GitHub: https://github.com/dbinky/Pommel

Still early (v0.3.x), but I figured the overlap in philosophy was worth mentioning. Both tools are trying to give agents memory/context without the token cost.

Happy to answer questions if anyone's curious.

[davidcforbes]

Is this similar to Serena, which uses LSP and Semantec search? That is what I use, and it is pretty good.

[dbinky]

I'll have to check it out!

[peterkc]

@davidcforbes I've been having issues with Serena and large repos. The MCP is rather bloated in context usage. Any setup tips?

I'm currently using cli-lsp-client, which uses hooks and CLI.

[dbinky]

Part of the point of tools like Beads is that MCP servers tend to hog up context, while CLIs are fast and context-efficient. That's what part of beads inspired me to create Pommel.

[davidcforbes]

Just a note on the serena LSP mcp memory footprint: On Windows it appears to be 29 tools and about 7500 tokens:
└ mcp__plugin_serena_serena__read_file: 339 tokens
└ mcp__plugin_serena_serena__create_text_file: 153 tokens
└ mcp__plugin_serena_serena__list_dir: 299 tokens
└ mcp__plugin_serena_serena__find_file: 184 tokens
└ mcp__plugin_serena_serena__replace_content: 638 tokens
└ mcp__plugin_serena_serena__search_for_pattern: 1.2k tokens
└ mcp__plugin_serena_serena__get_symbols_overview: 311 tokens
└ mcp__plugin_serena_serena__find_symbol: 1.1k tokens
└ mcp__plugin_serena_serena__find_referencing_symbols: 364 tokens
└ mcp__plugin_serena_serena__replace_symbol_body: 304 tokens
└ mcp__plugin_serena_serena__insert_after_symbol: 243 tokens
└ mcp__plugin_serena_serena__insert_before_symbol: 250 tokens
└ mcp__plugin_serena_serena__rename_symbol: 251 tokens
└ mcp__plugin_serena_serena__write_memory: 170 tokens
└ mcp__plugin_serena_serena__read_memory: 176 tokens
└ mcp__plugin_serena_serena__list_memories: 75 tokens
└ mcp__plugin_serena_serena__delete_memory: 130 tokens
└ mcp__plugin_serena_serena__edit_memory: 295 tokens
└ mcp__plugin_serena_serena__execute_shell_command: 364 tokens
└ mcp__plugin_serena_serena__activate_project: 109 tokens
└ mcp__plugin_serena_serena__switch_modes: 130 tokens
└ mcp__plugin_serena_serena__get_current_config: 81 tokens
└ mcp__plugin_serena_serena__check_onboarding_performed: 95 tokens
└ mcp__plugin_serena_serena__onboarding: 94 tokens
└ mcp__plugin_serena_serena__think_about_collected_information: 124 tokens
└ mcp__plugin_serena_serena__think_about_task_adherence: 124 tokens
└ mcp__plugin_serena_serena__think_about_whether_you_are_done: 89 tokens
└ mcp__plugin_serena_serena__prepare_for_new_conversation: 80 tokens
└ mcp__plugin_serena_serena__initial_instructions: 124 tokens

[davidcforbes]

If Pommel can work across the entire hard drive, then it would be a benefit, as Serena creates separate indexes in each project repository. So it doesn't help when your agent needs to search across the entire hard drive.

[davidcforbes]

I absolutely agree with the desire to use well-cratfed CLIs instead of MCP. I think ripgrep was a good attempt, and Anthropic's adoption of LSP, is another early attempt to build upon existing tooling. With that said, both of the CLIs has issues, so if we need to design and build a proper cli search tool, I am game ;-)

[dbinky]

That's what I tried to do with Pommel, but code-centric with semantic vector search.

[peterkc]

@dbinky would love to contribute. Will take it for a test run and submit PRs.