AgentsView: see what your AI coding agents did and what it cost

AgentsView brings your AI coding sessions into one searchable archive. Browse recorded conversations from more than 60 agent formats, compare activity and costs, and reuse lessons from past work. The archive stays on your machine unless you turn on a feature that shares it.
macOS and Linux · Windows, desktop app, pip, and Docker installs
Local-first · MIT · Single Go binary · Desktop · Web · CLI · MCP · SQLite / PostgreSQL / DuckDB / ClickHouse
A background server watches the session directories your agents already write and imports supported formats into a local SQLite archive. It finds the default directories automatically. You can configure other locations. Start a session in one window and watch it stream into AgentsView in another.
Import Claude.ai chats through Chrome, including Google sign-in. Enable Codex Cloud on one machine to fetch ChatGPT dot, Codex web, and cloud task chats. Cloud chats without token counts show plan usage; optional live capture records only turns heard while connected and can send queued messages.
The Activity dashboard shows peak concurrency and the exact moment it happened, active versus idle time, agent-minutes across parallel sessions, and cost. Scope it to any day, week, month, or custom range, and filter by project, agent, and machine. Switch from concurrency to user-message or assistant-message frequency to compare prompts and output. The Analytics dashboard adds message-role comparisons and Git outcome totals.
Token and cost reports read from the pre-indexed archive instead of reparsing raw session files every time. Pricing tracks LiteLLM and OpenRouter rates with an offline fallback, and cache-aware accounting covers prompt-cache creation and reads. Compare Total Input with Uncached Input, and choose Smooth, Lines, or Bars for the daily cost chart. Select a cost breakdown item to narrow totals and charts, open hermes-cron to compare jobs, and group model variants with configurable aliases.
Find a conversation by project and date in the command palette, or open it by ID with Ctrl/Cmd+G. Full-text search matches words; search by meaning finds related ideas, and you can turn it on from the palette. Health scores and quality signals link back to the transcript so you can check the evidence. Tool sequences follow errors, retries, switches, and recovery through the recorded calls without running a model. Tool Usage rates link to sessions with empty results, repeated calls, and recovery. For an assessment of one session, Analyze tool effectiveness generates a report with citations to the calls it judged. The optional Friction Log reviews each day's corrections, tool errors, and recurring problems, and can file them as Kata issues.
Recall (experimental) extracts decisions, warnings, and project facts from your archive. Each entry links to the messages that support it. Approve useful entries or archive rejected ones; your decisions survive later extraction changes. Generated Insights write model-authored reports over an explicit session scope.
The same archive is available to your agents through a CLI, a REST API, and an MCP server. An agent can check what a previous session already tried, quote its own history, or watch its spend mid-run. Read-only MCP tools expose file edits, tool calls, child sessions, Activity reports, tool usage, cost comparisons, and offline guide search. Native conversation-memory plugins bundle recall instructions and session-start refresh for Claude Code and Codex. See memory setup. Run agentsview why to find the session and prompt behind a line of code.
SSH sync fetches sessions from Linux and macOS hosts without a manual AgentsView install or a remote daemon. Local archives preserve a retired computer's history, annotations, and original Claude and Codex files.
SQLite is the local archive of record. From there, push parsed sessions to PostgreSQL or ClickHouse, mirror into DuckDB, read provider files from S3, or keep original source files in hosted storage.
An installable auto-push service mirrors each machine's archive into PostgreSQL; a read-only server presents the merged fleet. PostgreSQL sync.
Mirror into DuckDB for local or Quack reads, or push to ClickHouse and serve the dashboard there, including semantic and hybrid search when embeddings are pushed. Keep session edits in SQLite.
Point a primary viewer at out-of-band copies of native session directories, including S3-backed sources. Filesystem sync.
Upload original session files to an operator-managed server with device credentials and resumable transfers. Opt-in server parsing makes those sessions browsable in PostgreSQL without a local SQLite archive or a separate pg push. Hosted raw sync.
Loopback by default, with explicit flags for SSH forwards, remote dev environments, and authenticated non-loopback exposure. Remote access.
Move immutable session revisions between trusted folders for experimental portable workflows. Artifact folder sync.
Your agent transcripts can contain source code, credentials, paths, and private conversations. AgentsView starts with one local SQLite archive and a loopback-only server. Session content leaves the machine only through features you choose.
Browse, search, report, and run local Recall workflows without sending session content to an AgentsView account.
Configure PostgreSQL, ClickHouse, remote DuckDB, Generated Insights, GitHub publishing, or hosted raw sync only when that outcome is worth sending data.
Each sharing feature documents what it sends, where it goes, and which controls apply. Privacy details.
Install AgentsView and it finds the sessions that are already on your machine. The guide walks the whole loop in five minutes; the docs cover installation, workflows, and exact command behavior.
Copyright 2026 Kenn Software LLC. Open source under MIT.
Documentation capture