A token-frugal MCP server that gives coding agents structural understanding of a codebase: repo maps, symbol outlines, ranked search, import tracing, and parsed diagnostics — every result under a toke
View on GitHub ↗An MCP server that lets coding agents understand a codebase without reading it.
Repository maps, symbol outlines, ranked search, import tracing, change review, and parsed diagnostics — every result under a token budget it actually honours.
Most MCP servers are thin wrappers around readFile and grep. That makes them easy to write and expensive to use. An agent asked to add a field to a service typically burns through something like:
Roughly 14,000 tokens to learn four facts. The context window fills with material the model will never refer to again, and the useful signal is buried inside it.
Jelly Bean attacks this directly. Same task, same repository:
Under 900 tokens, and the agent knows strictly more than it did above — because it also knows what else is in those files, and what depends on them.
Everything here follows from three decisions.
1. Handles instead of dumps. Every result is compact rows carrying an identifier like jb7c1e9a04 that addresses an exact region of an exact file. The agent spends tokens on a body only for the handles it chooses to follow. Handles are content-derived, so the same region always yields the same handle — an agent can tell that a search hit and an outline entry are the same function without expanding either.
2. Budgets are a contract, not a hint. Every tool takes tokenBudget and fits inside it, degrading detail rather than truncating mid-row. The footer always reports what was omitted and names the one call that would reveal it, so a truncated result is a signpost instead of a dead end.
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/rayhankhilji--jellybean.html)
Multi-agent orchestration for Claude Code: MCP server + skill that keeps Claude as lead engineer while delegating backend work to OpenAI Codex and frontend/design work to Kimi
herry2059/project-os-for-codexOpen-source control plane for Codex projects: Git-backed context, visible agent progress, scoped MCP access, resumable work, and safe handoffs.
anshupriyan/cli-bridgeMCP server for Claude Desktop and other Model Context Protocol clients — sandboxed filesystem access (read/write/edit/search) plus Dev Mode-gated shell command execution, scoped to a single workspace
roshanarnav25-sloth/google-flow-mcpMCP server that drives Google Flow (Veo / Nano Banana) through its internal tRPC API, with a hard credit-cost gate
Gowthaman90/mcp-bastionReliability + security proxy for the Model Context Protocol (MCP): self-healing connections, runtime tool-security (rug-pull/poisoning detection), and a compliance-mapped audit trail.
egorfedorov/mozgExam-scored knowledge brains for AI agents: paste a docs URL, get a searchable brain over MCP with a measured score and known gaps. AGPL.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.