agentsview
Browse, search, and track costs across all your AI coding agents. One binary, no
accounts, everything local.
Install
macOS and Linux:
curl -fsSL https://agentsview.io/install.sh | bash
Windows (PowerShell):
powershell -ExecutionPolicy ByPass -c "irm https://agentsview.io/install.ps1 | iex"
Or download the desktop app (macOS / Windows) from
GitHub Releases or via
homebrew: brew install --cask agentsview
Or run the published Docker image:
docker run --rm -p 127.0.0.1:8080:8080 \
-v agentsview-data:/data \
-v "$HOME/.claude/projects:/agents/claude:ro" \
-v "$HOME/.forge:/agents/forge:ro" \
-e CLAUDE_PROJECTS_DIR=/agents/claude \
-e FORGE_DIR=/agents/forge \
ghcr.io/kenn-io/agentsview:latest
Quick Start
agentsview serve # start the server in the foreground
agentsview daemon start # alternatively, start the same server in the background
agentsview daemon status # show daemon status
agentsview daemon restart # restart from current configuration
agentsview daemon stop # stop the server, web UI, and sync
agentsview session list # read from the daemon if warm, otherwise SQLite
agentsview usage daily # print daily cost summary
On first run, agentsview discovers sessions from every supported agent on your
machine, syncs them into a local SQLite database, and serves a web UI at
http://127.0.0.1:8080.
The daemon is the server: the web UI, API, session sync, and file watchers share
one process. Choose serve for a foreground process or daemon start for a
background process. Running serve when a compatible daemon is already running
reports its URL and exits. daemon stop and serve stop both stop that
writable server, including sync; serve stop also stops read-only mirror
servers for the same data directory.
For Devin CLI, point DEVIN_DIR or agents.devin.dirs at the local root that
contains cli/ — for example ~/Library/Application Support/devin on macOS,
~/.local/share/devin on Linux, or a redacted path like
.../Application Support/devin. AgentsView reads session data under
<root>/cli/... and intentionally ignores copied config or OAuth paths. Do not
paste tokens, OAuth files, or other secrets into bug reports.
Claude and Codex sources can also be configured as s3:// roots, so a central
AgentsView instance can read sessions that other machines push to S3-compatible
object storage. Add those roots to agents.claude.dirs or agents.codex.dirs;
AgentsView lists object metadata and only downloads changed sessions during
sync. S3 change detection uses size, modified time, and available object
fingerprints such as ETag, version ID, or checksums.
The desktop app and freshness-sensitive CLI commands share a detached local
daemon. Read-only CLI commands attach to it when it is already running, but fall
back to direct read-only SQLite on a cold archive so one-off scripts stay fast.
Commands that need fresh data or need to write, such as sync, usage,
token-use, pg push, and duckdb push, auto-start the daemon when needed.
The server remains running after these commands exit and also serves the web UI.
For a one-shot sync with no background server, stop the daemon first and run
AGENTSVIEW_NO_DAEMON=1 agentsview sync.
Use agentsview daemon start when you want to start the writable SQLite daemon
explicitly. It loads the normal effective configuration from config.toml and
supported environment variables; daemon start and daemon restart accept no
serve-specific flags. Background daemons self-exit after an idle period unless a
client request or daemon-owned job is active.
The existing agentsview serve --background, agentsview serve status, and
agentsview serve stop commands remain available. Use serve --background when
a one-off daemon needs a serve-only flag, such as --no-sync or an
unauthenticated non-loopback --host override.
Remote / forwarded access
agentsview binds to loopback and validates the request Host header to guard
against DNS-rebinding attacks. When you reach it through SSH port-forwarding, a
reverse proxy, or a remote dev environment (exe.dev, Codespaces, Coder, WSL2),
the browser sends a Host that the server does not recognize, so API requests
such as /api/v1/settings are rejected with 403 Forbidden.
To fix this, restart the server with --public-url set to the exact origin you
open in the browser:
# Browser opens http://127.0.0.1:18080 via `ssh -L 18080:127.0.0.1:8080 host`
agentsview serve --public-url http://127.0.0.1:18080
# Browser opens a forwarded hostname
agentsview serve --public-url https://your-workspace.exe.dev
--public-url also selects the URL opened on startup; it does not configure a
listener or external proxy. Use --no-browser on a headless server, or use
--public-origin (repeatable or comma-separated) on its own for trust-only
configuration. If you expose the UI beyond loopback, also enable
--require-auth. See
Remote Access
for the listener, browser URL, and managed Caddy port rules.
Docker
The container image defaults to local agentsview serve. Set PG_SERVE=1 to
switch the startup command to agentsview pg serve instead.
docker-compose.prod.yaml is included as a production example:
docker compose -f docker-compose.prod.yaml up -d
The included compose file persists the agentsview data directory in a named
volume and mounts Claude, Codex, Forge, and OpenCode session roots read-only.
The container runs as root, so prefer a named volume for /data over a host
bind mount; if you do bind-mount, pre-create the directory with the desired
ownership to avoid root-owned files in your home directory.
The examples publish the UI on loopback only (127.0.0.1). If you need to
expose it beyond localhost, enable --require-auth and publish the port
intentionally.
Important: a containerized agentsview instance can only discover agent sessions
from directories you explicitly mount into the container. If you do not mount an
agent's session directory and point the matching env var at it, that agent will
not appear in the UI.
Example PostgreSQL-backed startup:
docker run --rm -p 127.0.0.1:8080:8080 \
-e PG_SERVE=1 \
-e AGENTSVIEW_PG_URL='postgres://user:[email protected]:5432/agentsview?sslmode=require' \
ghcr.io/kenn-io/agentsview:latest
Example DuckDB mirror startup:
# Populate /data/sessions.duckdb from the mounted SQLite archive.
docker run --rm \
-v agentsview-data:/data \
-v "$HOME/.claude/projects:/agents/claude:ro" \
-e CLAUDE_PROJECTS_DIR=/agents/claude \
ghcr.io/kenn-io/agentsview:latest duckdb push --full
# Serve the populated mirror read-only.
docker run --rm -p 127.0.0.1:8080:8080 \
-v agentsview-data:/data \
ghcr.io/kenn-io/agentsview:latest duckdb serve
Example Quack startup:
# Expose the local DuckDB mirror over Quack from the host/container.
QUACK_TOKEN="$(openssl rand -base64 32)"
docker run --rm -p 127.0.0.1:9494:9494 \
-v agentsview-data:/data \
ghcr.io/kenn-io/agentsview:latest \
duckdb quack serve \
--bind quack:0.0.0.0:9494 \
--token "$QUACK_TOKEN" \
--allow-insecure
# Serve the web UI from a remote Quack endpoint.
# The client url uses the native quack:HOST:PORT form; the Quack extension
# does not accept quack:http:// or quack:https:// urls. A non-loopback host
# needs --allow-insecure.
docker run --rm -p 127.0.0.1:8080:8080 \
-e AGENTSVIEW_DUCKDB_URL='quack:duckdb.example.com:9494' \
-e AGENTSVIEW_DUCKDB_TOKEN="$QUACK_TOKEN" \
ghcr.io/kenn-io/agentsview:latest duckdb serve --allow-insecure
Keep Quack on loopback or behind TLS. Plain HTTP Quack on a non-loopback bind
requires --allow-insecure and should only be used behind a trusted tunnel or
reverse proxy.
Token Usage and Cost Tracking
agentsview usage tracks token consumption and compute costs across all
your coding agents -- not just Claude Code. Reports read from the same local
SQLite archive that powers the UI.
# Daily cost summary (default: last 30 days)
agentsview usage daily
# Per-model breakdown
agentsview usage daily --breakdown
# Filter by agent and date range
agentsview usage daily --agent claude --since 2026-04-01
# One-line summary for shell prompts / status bars
agentsview usage daily --all --json
agentsview usage statusline
Features:
- Automatic pricing via LiteLLM and OpenRouter rates (with offline fallback)
- Authoritative Copilot CLI billing totals when session logs provide them
- Prompt-caching-aware cost calculation (cache creation / read tokens)
- Per-model breakdown with
--breakdown
- Date filtering (
--since, --until, --all), agent filtering (--agent)
- JSON output (
--json) for scripting
- Timezone-aware date bucketing (
--timezone)
- Works standalone -- no server required, just run the command
Per-Session Details
agentsview session usage <id> prints per-session token statistics plus a cost
estimate for a single session. The output reports the session's total output
tokens and peak context tokens, plus a cost estimate (cost) when pricing is
available for the session's model(s) (has_cost). Cost is computed from
input/output and cache tokens internally, but only the output-token and
peak-context totals are reported alongside the cost.
# Print token usage and cost for a specific session
agentsview session usage <id>
# JSON output for scripting
agentsview session usage <id> --format json
The same per-session usage data is available from the REST API:
GET /api/v1/sessions/{id}/usage
The response includes the session_id, agent, project,
total_output_tokens, peak_context_tokens, has_token_data, cost,
has_cost, models, and unpriced_models fields from the CLI JSON schema.
Machine-readable money is always an integer microdollar object, for example
{"cost":{"microdollars":2410000}}; CLI tables and labels render that value as
ordinary dollars. HTTP responses also include server_running: true. Existing
sessions return 200 even when token or cost data is absent; missing sessions
return 404.
The deprecated alias agentsview token-use <id> remains available for
compatibility and now also reports cost estimates.
For one exact non-interactive claude -p or codex exec --json execution in
CI, use agentsview capture run. It preserves the child streams and exit
outcome, writes a separate versioned usage result, and starts no daemon, web
server, or watcher. See
One-shot CI capture.
Session Stats
agentsview stats emits window-scoped analytics over recorded sessions: totals,
archetypes (automation vs. quick/standard/deep/marathon), distributions for
session duration, user-message count, peak context, and tools-per-turn, plus
cache economics, tool/model/agent mix, and a temporal hourly breakdown. The
--format json output follows a versioned v1 schema (schema_version: 1)
suitable for downstream consumers.
By default, stats only reads the local SQLite archive. Git-derived outcome
metrics are opt-in because they can be slow or brittle on large/missing repos:
use --include-git-outcomes for commits/LOC/files changed, and
--include-github-outcomes for GitHub PR counts via gh (this also enables git
outcomes).
# Human-readable summary over the last 28 days
agentsview stats
# Machine-readable JSON over a fixed date range
agentsview stats --format json --since 2026-04-01 --until 2026-04-15
# Restrict to one agent and inspect the schema
agentsview stats --format json --agent claude | jq '.schema_version'
# Include expensive local git outcome metrics explicitly
agentsview stats --include-git-outcomes
Session Browser
| Dashboard |
Session viewer |
|
 |
| Search |
Activity heatmap |
 |
 |
- Full-text search across all message content (FTS5)
- Semantic search (opt-in) -- index session content with any
OpenAI-compatible embeddings endpoint and search by meaning with
agentsview session search --semantic or --hybrid; every content-search
match cites the conversation unit it came from
(docs)
- Token usage and cost dashboard -- per-session and per-model cost
breakdowns, daily spend charts, all in the web UI
- Analytics dashboard -- activity heatmaps, tool usage, velocity metrics,
project breakdowns
- Recent Edits feed -- the files your agents changed most recently across
every session, grouped by project and path, each linking to the message that
made the change
- Data workspace -- inspect project inventory and observed folders, preview
reclassification impact, and manage worktree mapping rules
- Recall corpus browser -- explore experimental distilled knowledge and jump
from entries to their supporting transcript evidence
- Live updates via SSE as active sessions receive new messages
- Keyboard-first navigation (
j/k/[/], Cmd+K search, ? for all
shortcuts)
- Export sessions as HTML or publish to GitHub Gist
Supported Agents
agentsview discovers sessions from all of these. Aider is opt-in because it has
no central session directory; set AIDER_DIR or agents.aider.dirs to enable
it. Amp support is deprecated because current Amp releases may store threads
server-side and leave only local stubs; agentsview can still parse historical
local Amp thread JSON files.
| Agent |
Session Directory |
| Aider |
<repo>/.aider.chat.history.md (per repo; opt in with AIDER_DIR or agents.aider.dirs) |
| Amp (deprecated) |
~/.local/share/amp/threads/ (historical local thread JSON only) |
| Antigravity |
~/.gemini/antigravity/ |
| Antigravity CLI |
~/.gemini/antigravity-cli/ (see note below) |
| Claude Code |
~/.claude/projects/ |
| OpenClaude |
~/.openclaude/projects/ |
| Claude Cowork |
~/Library/Application Support/Claude/local-agent-mode-sessions/ (macOS) |
| Codex |
~/.codex/sessions/ |
| Copilot CLI |
~/.copilot/ |
| Devin CLI |
~/.local/share/devin/ (Linux), ~/Library/Application Support/devin/ (macOS); point DEVIN_DIR / agents.devin.dirs at the root that contains cli/ |
| Evener |
~/.local/state/evener/projects/ (or $XDG_STATE_HOME/evener/projects/; semantic transcript v2) |
| Cortex Code |
~/.snowflake/cortex/conversations/ |
| Cursor |
~/.cursor/projects/ |
| DeepSeek TUI |
~/.codewhale/sessions/, ~/.deepseek/sessions/ |
| DeepSeek Harness |
~/.dsh/sessions/ (or $DSH_HOME/sessions/) |
| Forge |
~/.forge/ |
| Gemini CLI |
~/.gemini/ |
| Goose |
~/.local/share/goose/sessions/ (macOS a |
AgentsView 是一款本地优先的工具，用于浏览、搜索和追踪 Claude Code、Codex 以及其他 20 多种编码智能体会话的费用。所有数据都保存在本地 SQLite 数据库中，无需任何云端账户。