Map

agentsview

Toolbox toolboxgosvelteai-agentsanalyticscost-trackingsqlitelocal-firstkenn-softwareGoMIT ↳ show in map Markdown
title
agentsview
type
toolbox
summary
Local-first browser, search and cost tracker for sessions across 50+ coding agents
tags
go, svelte, ai-agents, analytics, cost-tracking, sqlite, local-first, kenn-software
repo
https://github.com/kenn-io/agentsview
language
Go
license
MIT
sources
agentsview
created
2026-05-22
updated
2026-07-21

Local-first session intelligence for coding agents: browse, full-text search, and cost-track everything you've done across Claude Code, Codex, Copilot CLI, Cursor, OpenHands, forge, and 50+ other agents — in one single binary, no account, all data on your machine. Doubles as a fast ccusage replacement. Part of the kenn-software-suite and the most mature member of it (~4,500 stars).

What it does

Three things, all over your local agent session files.

Search and browse. Full-text search (SQLite FTS5) across all message content by default, with optional semantic and hybrid search when you point it at any OpenAI-compatible embeddings endpoint. Live updates over SSE for active sessions, a cost dashboard, activity heatmaps, tool-usage metrics, and a Recent Edits feed. Sessions export to HTML or publish as a GitHub Gist.

Cost tracking (agentsview usage). Billed as "a fast, local replacement for ccusage and similar tools." Automatic pricing from LiteLLM rates, prompt-caching-aware cost math, per-model and per-session breakdowns, date filtering, JSON output for scripting. Pricing works offline with a bundled fallback table when the rate feed is unreachable. The speed claim: token-usage queries are "100x faster than re-parsing session files" because agentsview pre-indexes sessions into SQLite instead of re-reading JSONL each time.

Stats (agentsview stats, versioned v1). Window-scoped analytics — totals, session archetypes (automation, quick, standard, deep, marathon), duration distributions, context metrics, cache economics, hourly temporal breakdowns, and opt-in Git/GitHub outcome metrics (commits, LOC, files changed, merged PRs).

How it works technically

Go 1.26+ backend with a Svelte 5 SPA (Vite/TypeScript) frontend and a Tauri desktop wrapper. Storage is tiered: local SQLite with FTS5 by default, an optional PostgreSQL backend for team dashboards, and a DuckDB mirror for analytical queries. Layout:

cmd/agentsview/   CLI entrypoint
internal/         config, db, parser, server, sync, postgres
frontend/         Svelte 5 SPA
desktop/          Tauri wrapper

A detached local daemon does the indexing and is shared by both the desktop app and the CLI, so they read one warm index instead of each re-parsing; it self-exits when idle. Session sources aren't limited to local directories — Claude and Codex roots can be s3:// prefixes, with ETag/checksum change detection so only changed session files get re-pulled.

Design choices that recur across the kenn-software-suite: local-first with no telemetry (the server binds loopback by default, and telemetry is opt-out via AGENTSVIEW_TELEMETRY_ENABLED=0), speed via indexing, keyboard-first UI (j/k/[/], Cmd+K search, ? for shortcuts). Expensive metrics are opt-in — Git outcomes and GitHub PR counts are off by default because they're slow or brittle.

Multi-user teams push local session data to a shared Postgres for team dashboards via agentsview pg push and a PG_SERVE=1 serve mode. The DuckDB mirror can also be exposed over a network endpoint using the Quack wire protocol for external analytical clients. Exposing any of this remotely is gated: the server validates the Host header against DNS-rebinding attacks, and remote access requires explicit --public-url and --require-auth flags rather than being on by default.

Basic usage

curl -fsSL https://agentsview.io/install.sh | bash
agentsview serve            # start server, open the web UI
agentsview usage daily      # daily cost summary
agentsview stats            # window-scoped session analytics

Each agent's session directory is configurable via environment variables. Build needs Go 1.26+ with CGO enabled and Node 22+.

Limitations

  • Containerized instances only see agent sessions from directories explicitly mounted into the container.
  • Git outcome metrics need local repo access; GitHub outcome metrics need the gh CLI.
  • ctx — local SQLite context manager for Claude Code/Codex; ctx organizes/binds conversations, agentsview analyzes and costs them.
  • gomodel / routing-run — LLM gateways where the spend happens; agentsview is the cost ledger after the fact.
  • roborev — sibling tool that scores the quality of agent output, where agentsview scores the volume and cost.
  • kanbots — dispatches and runs the agents whose sessions agentsview then indexes and prices.
  • duckdb-quack-protocol — the wire protocol agentsview's DuckDB mirror speaks over the network for external analytical clients.
  • kenn-software-suite — the rest of the suite.

Repo

github.com/kenn-io/agentsview — Go (1.26+) + Svelte 5 + Tauri, MIT, ~4,500 stars. agentsview.io. No telemetry, no analytics, no accounts.

Linked from