Track AI coding agent sessions and costs in a local web interface
Using multiple AI coding tools simultaneously makes it difficult to track which session used which tokens or which files were modified. agentsview collects locally stored session data to allow searching and analysis through a web interface. The latest version, 0.46.1, fixes Antigravity transcript issues and read-only mount problems, preserves repeated prompts, and restores container builds.
Local data-based integrated archive
agentsview automatically discovers AI agent session directories on the user's machine and syncs them to a local SQLite database. During this process, the session archive remains local unless the user selects specific features. The server handles the web UI, API, session synchronization, and file watcher in a single process, which can be run in the foreground with the serve command or in the background with daemon start. On first run, it discovers sessions for all supported agents and serves the web UI at http://127.0.0.1:8080. When accessing via SSH port forwarding or a reverse proxy, the --public-url flag must be used to specify the exact origin opened in the browser to avoid 403 errors.
Support for over 50 agents and session exploration
It supports sessions from more than 50 AI coding agents, including Claude Code, Codex, Cursor, and Windsurf. Aider requires setting the AIDER_DIR environment variable because it lacks a central session directory, while Amp remains in a non-recommended state because the current release can store threads on the server side. The web UI provides full-text search, semantic search, token usage dashboards, activity heatmaps, and tool usage frequency. The 'Recent Edits' feed specifically groups files recently changed by agents by project and path, linking each item to the message that created the change. It also supports eight languages and updates new messages from active sessions in real time via SSE.
Token cost tracking and analysis reports
The agentsview usage command tracks token consumption and compute costs for all coding agents. It automatically applies rates from LiteLLM and OpenRouter and supports offline fallbacks. It provides cost calculations considering prompt caching, detailed analysis by model, and filtering by date and agent. The agentsview stats command outputs window-range analysis including distributions of session duration, user message count, peak context, and tools per turn. Git-based outcome metrics can be slow or unstable, so they must be explicitly enabled using the --include-git-outcomes flag. All reports run independently without starting the server and can be used for scripting via JSON output.
Installation and deployment methods
On Windows, run irm https://agentsview.io/install.ps1 | iex in PowerShell or download the desktop app from GitHub Releases. If using Homebrew, the brew install --cask agentsview command is also supported. It includes examples for integrating with various data sources such as PostgreSQL, DuckDB, and Quack, and production environment configuration can be referenced via docker-compose.prod.yaml.
Pre-use considerations
agentsview follows a local-first principle, but enabling semantic search sends data to an OpenAI-compatible embedding endpoint, so the scope of data sharing should be verified when using this feature. In Docker deployments, agent session directories not mounted to the container will not appear in the UI, so session paths for required agents must be explicitly mounted and environment variables set. Agents without a central repository, such as Aider, require limiting the directory scan scope to avoid macOS privacy prompts. Additionally, remote access via SSH will reject API requests if the --public-url setting is missing, so the appropriate flag for the network configuration must always be specified.