Background: from skeptic to all-in on AI
Wes McKinney, the creator of pandas and a foundational figure in Python data science, was reportedly a longtime AI skeptic — at one point blocking all LLM-related topics on Twitter. That changed in late 2025, when he left his stable job, founded Kenn Software, and began shipping multiple AI products in five months. His latest is agentsview, described as:
> Browse, search, and track costs across all your AI coding agents. One binary, no accounts, everything local.
McKinney himself is an extreme test case: he reportedly runs 27 different AI coding agents simultaneously — Claude Code, Codex, Cursor, OpenCode, Kimi, Qwen Code, OpenClaw, and more. Unable to find past sessions across dozens of tool-specific log formats, he concluded the problem was infrastructure, not tooling.
What agentsview does
- Unified cost tracking (
agentsview usage daily --breakdown,--agent,--since,statusline): indexes all sessions into SQLite once, making queries 100x+ faster than ccusage, which re-parses raw files on each run. Pricing comes from LiteLLM rates with offline fallback, and it correctly accounts for prompt caching (cache-creation vs. cache-read token pricing). - Full-text search: SQLite FTS5 indexing across all agent messages, with a Web UI (
agentsview serveathttp://127.0.0.1:8080) featuring token cost dashboards, activity heatmaps, and SSE live updates. Keyboard-first:j/knavigation,[/]session switching,Cmd+Ksearch,?for shortcuts. - Team sharing:
agentsview pg push/pg serve/pg push --watchsyncs SQLite data to a read-only PostgreSQL backend for managers, engineers, and audit use cases, with an optional launchd/systemd daemon. - Stack: Go 1.26+ backend (CGO for SQLite FTS5), Svelte 5 + Vite + TypeScript frontend, Tauri desktop wrapper.
- Three backends: SQLite (primary local store), PostgreSQL (optional team sync), DuckDB (optional mirror/analytics, remotely accessible via the Quack protocol).
- Local-first security: server binds to
127.0.0.1by default, host header validation against DNS rebinding,--public-urlfor SSH forwarding,--require-authwhen exposed, no data upload, optional anonymous-heartbeat telemetry only. - Encrypted sessions: for Antigravity CLI's encrypted
.pbfiles, a sidecar approach (agy-reader) decrypts locally and emits.trajectory.jsonfiles that agentsview reads — decryption never leaves the user's machine. - Developers have moved from picking one agent to running an agent matrix; session management becomes an exponential mess at team scale.
- Cost transparency is the last mile of enterprise AI adoption — agentsview normalizes everything into tokens + dollars.
- Data ownership: where competitors store sessions in vendor clouds, agentsview keeps everything local, treating session data as user assets.
- macOS/Linux only (Docker/WSL for Windows)
- Keyword-based FTS5 search, no semantic search
- Dashboards but no LLM-powered summarization
- No plugins or IDE integrations yet
- agentsview GitHub: https://github.com/kenn-io/agentsview
- Project site: https://www.agentsview.io/
- Wes McKinney: https://github.com/wesm
- Kenn Software: https://kenn.io/
Supported agents (27)
Claude Code (~/.claude/projects/), Codex, Copilot CLI, Gemini CLI, OpenCode, OpenHands CLI, Cursor, Amp, iFlow, Zencoder, Zed, VSCode Copilot, Pi, Qwen Code, OpenClaw, QClaw, Kimi, Kiro CLI/IDE, Cortex Code, Hermes Agent, WorkBuddy, Forge, Piebald, Warp, Positron Assistant, and Antigravity — coverage extends beyond mainstream tools into the long tail.
Architecture and design
Ecosystem and maturity
Early adopters include Simon Willison (whose claude-code-transcripts inspired the project) and Andy Fischer (claude-history-tool). The project sits at ~1,648 GitHub stars and version 0.1.x: early-stage risk, but high engineering credibility given McKinney's track record.
Why it matters
Limitations
Likely evolution: embedding-based semantic search, automatic session summaries, cross-agent session correlation, and cost-optimization recommendations.