Agent Analytics and Utilities
Observability, token optimization, and code indexing tools for AI coding agent workflows.
Parent: Agentic Development Tools
Agentlytics — Session Analytics Dashboard
| Repo | github.com/f/agentlytics (Fatih Kadir Akin) |
| Language | JavaScript (Node.js backend) + React 19 SPA (Vite + Tailwind) |
| What | Read-only analytics dashboard for AI coding sessions across 16 editors |
| Maturity | 167 commits in 5 days, v0.2.8, macOS only, no tests |
Reads session data from Cursor, Windsurf, Claude Code, VS Code Copilot, Zed, Gemini CLI, Codex, and more by parsing their local storage files (JSONL, SQLite, ConnectRPC). Normalizes into a local SQLite cache, serves a web dashboard at localhost:4637.
Features: KPIs, cost estimation, token usage, model/tool breakdowns, activity heatmaps, coding streaks, subscription/quota tracking, per-project analytics, editor comparison, SQL viewer for direct cache queries.
Relevance: useful for understanding token costs across all your editors. The subscription tracking (reads Keychain auth tokens to query Anthropic/Cursor/Copilot APIs) could help manage the limited-tokens home setup.
Distill — Token Compression Pipe Filter
| Repo | github.com/samuelfaj/distill |
| Language | TypeScript/Bun, distributed as native platform binaries via npm |
| What | Pipe verbose command output through a local LLM to compress it before feeding back to an expensive model |
| Maturity | 29 commits over 6 days, v1.4.1 |
bun test 2>&1 | distill "Did the tests pass?" — claims up to 99% token savings. Default backend: Ollama with qwen3.5:2b. Supports 11 LLM providers.
Modes: batch (one-shot summarization), watch (auto-detects repeating output, reports diffs), interactive passthrough (detects [y/N] prompts). Bad distillation detection falls back to raw output.
Relevance: useful at home with limited tokens. Each Claude Code session could pipe verbose outputs through distill. Requires a local model (Ollama).
jcodemunch-mcp — Token-Efficient Code Indexing
| Repo | github.com/jgravelle/jcodemunch-mcp |
| Language | Python 3.10+ |
| What | MCP server that indexes codebases with tree-sitter, provides O(1) symbol retrieval instead of loading entire files |
| Maturity | 245 commits, v1.5.1, 5+ contributors, dual license (free personal / paid commercial) |
Instead of an agent reading 40,000 tokens to find one function, it reads ~200 tokens via search_symbols then get_symbol. Supports 30+ languages via tree-sitter.
Features: 14+ MCP tools, incremental re-indexing (hash-based change detection), stable symbol IDs, multiple transports (stdio, SSE, HTTP), bearer token auth for HTTP mode.
Relevance: direct token saver for Claude Code sessions. Configure as MCP server per project. The 30+ language support covers Go, Rust, Python, TypeScript — all relevant to your work.
Open Terminal — Remote Terminal API
| Repo | github.com/open-webui/open-terminal (Open WebUI org) |
| Language | Python 3.11+ (FastAPI + Uvicorn) |
| What | Headless API server that gives AI agents a sandboxed environment for shell commands, file management, Jupyter execution |
| Maturity | 128 commits, v0.11.20, by Timothy Jaeryang Baek (Open WebUI creator) |
REST API for command execution, file CRUD, WebSocket interactive terminals (xterm.js compatible), Jupyter notebook cell execution. MCP server mode exposes all endpoints. Multi-user isolation with per-user Linux accounts. Network egress filtering via dnsmasq + iptables.
Relevance: could provide sandboxed execution environments for agents. The network egress filtering is interesting given the sandbox constraints on your work Mac. However, it’s a complementary execution backend, not a session manager.
Quick Reference
| Tool | Purpose | Token savings | MCP | Effort to adopt |
|---|---|---|---|---|
| Agentlytics | Analytics dashboard | Indirect (visibility) | No | Low (just install + run) |
| Distill | Compress command output | Direct (up to 99%) | No | Low (pipe filter) |
| jcodemunch-mcp | Code indexing | Direct (~95% per lookup) | Yes (14 tools) | Medium (index per project) |
| Open Terminal | Sandboxed execution | None | Yes | High (server deployment) |