One archive for a scattered working life
Coding agents remember a great deal, but they rarely make that memory easy to use. Claude Code, Codex, Cursor, OpenCode, Copilot, and their competitors each write sessions in different folders and formats. Finding why a file changed last month often means guessing which tool did the work, locating its private data directory, and reading JSONL by hand. AgentsView discovers those records, normalizes them into SQLite, and presents one web interface and command line.
The result is more useful than a chat viewer. Full-text search covers message content. The dashboard shows activity, tool and model mix, projects, recent edits, context use, and estimated spending. A session page connects conversation, tool calls, and usage data. Live updates arrive through server-sent events, while HTML export provides a portable record. Optional semantic search can use any OpenAI-compatible embeddings endpoint and cites the conversation unit behind each match.
Breadth is the headline. The README names more than 40 supported sources, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Copilot variants, Aider, Goose, Kimi, Qwen Code, and Windsurf. A narrower viewer may be cleaner for one tool, but it cannot answer cross-agent questions about cost or past work.
Local-first is a sensible default
The normal path is one binary and no account. agentsview serve discovers supported folders, syncs sessions into a local SQLite database, and binds the UI to 127.0.0.1:8080. The detached daemon lets the desktop app and CLI share a warm archive, then exits after an idle period when no client or job needs it. Read-only commands can query SQLite directly if the daemon is cold. This is considerate desktop behavior.
Privacy still needs active thought. Agent transcripts can contain source code, terminal output, file paths, pasted credentials, and internal decisions. Full-text indexing puts that material into another high-value file. Keep database permissions tight, protect backups, and scan the archive before publishing an HTML file or GitHub Gist. Semantic search remains local only if its embedding endpoint is local. PostgreSQL push, S3 source roots, and remote DuckDB access deliberately create new data boundaries.
AgentsView sends an anonymous daemon_active event to PostHog at startup and every 24 hours. The documented payload includes app version, commit, operating system, CPU architecture, and a random installation ID, not prompts, projects, paths, accounts, or machine identity. Two environment variables disable it. That is restrained telemetry, but privacy-sensitive deployments should set the opt-out explicitly.
The first sync is where reality appears
Installation options are excellent. The project offers shell installers, Homebrew, Windows packages, macOS builds, Linux archives, AppImages, and a container. Release assets include x86-64 and Arm builds for the major platforms. Version 0.40.1 shipped on August 4, 2026. Most people can be looking at a dashboard within minutes.
Discovery is less uniform because the source applications are not uniform. Aider has no central store, so users must choose roots for a bounded repository scan. JetBrains Copilot requires a separate exporter and repeated exports for fresh data. Current Amp releases may leave only server-side threads, so AgentsView marks that parser deprecated and handles historical local JSON. Antigravity can fall back to a summary unless another local tool produces full transcript sidecars. Docker sees only directories explicitly mounted and configured.
Large or odd archives deserve a copy-first trial. Issue #1412 describes a Copilot orchestration session whose inlined subagent output created a JSONL line larger than 100 MB. The parser rejected it and the wider sync stopped progressing. Issue #1416 reports that encrypted Trae stores can cause repeated full reconciliations and about 300 percent CPU on the reporter's machine. The companion issue #1417 says there is no reliable configuration switch to exclude one discovered agent source. These reports concern version 0.40.1 and were still open on August 14.
Usage numbers are helpful estimates
The usage commands are a strong reason to install AgentsView even if the dashboard is secondary. Daily summaries can filter by agent and date, split by model, account for prompt-cache reads and writes, and emit JSON for a status line or script. Because sessions are already indexed, normal queries do not reread every raw transcript. Per-session output includes total output tokens, peak context, models, missing pricing, and a cost estimate represented as integer microdollars in machine-readable responses.
Do not confuse precision of representation with billing authority. Model names change, agent logs omit fields, providers revise prices, and some models remain unpriced. The latest release specifically fixed Kimi Work tool-step tracking and K2.6 pricing. AgentsView can be excellent for trends, comparisons, and catching unexpected usage. Reconcile material spending against the provider's bill.
Team backends add real administration
A team deployment needs transcript policy, database credentials, authentication, TLS, retention, and project exclusions. PostgreSQL supports shared dashboards, while DuckDB and Quack serve analytics mirrors. The Docker examples bind only to loopback, and the server validates host headers against DNS rebinding. Remote users must configure the public origin, while any non-loopback exposure should require authentication. Those conservative defaults are a genuine strength.
Fast-moving and worth trying
AgentsView was created in February 2026, so its scope has expanded very quickly. The repository was pushed on August 15, the latest stable release arrived eleven days earlier, and detailed parser issues were filed on August 14. Its 93 open count includes issues and pull requests. This is active maintenance, but the 0.x version and current sync failures make archival completeness something to verify, not assume.
For a developer using several agents, the payoff is immediate: old conversations become searchable, costs become comparable, and edits regain context. Start locally, disable sources you do not need when that feature is available, and keep the raw logs as the source of truth. AgentsView is already a very good lens over coding-agent history. It is not yet the only copy of that history you should trust.