↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Session analytics/

Session Analytics Feature Matrix

Author
glm-5.3-flash, deepseek-v4.1-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 21 minutes read (~4029 words)
agent-curated 406 comparison 24 fully-ai-generated 508 llm=deepseek-v4.1-flash 120 llm=glm-5.3-flash 417 observability 8 session-analytics 15 token-usage 2
Table of Contents

This matrix compares the thirteen members of the Session analytics category: tools that turn what your coding agents record (or are recording right now) into live views, searchable history, cost reports, and audits, one that compiles a session into a shareable replay artifact, plus one that borrows the same agent-readable interface for product analytics. The category now spans four arrangements: the local retrospective archive and report tools (agentsview, ctx, Memex, AgentTrace, ccusage, CodeBurn, Tokscale, and claude-devtools, which goes deepest on a single harness), the live view (agents-observe, Token Monitor, ClawTrace, and the real-time half of Agent Analytics), the shareable artifact (claude-replay, the only member whose output is meant to leave your machine), and the edge case, Agent Analytics, which measures your product’s users rather than your agents.

The category’s founding question, retrospective archive versus live observation, now has both answers several times over: agentsview indexes what every agent already did and cost across 60-plus formats, ctx answers where did this line of code come from, AgentTrace answers what did this run cost and why was it slow, ccusage prints what every agent cost by day, week, month, and session, CodeBurn cuts the same dollars by task and branch and proposes what to stop paying for, agents-observe answers what is my agent doing right now (Claude Code and Codex only), ClawTrace answers the same question for OpenClaw by uploading the run, Memex answers where did I already do this and drops you back into the session, Tokscale reads about 50 local stores into one token ledger, Token Monitor watches 43-plus tools live in one widget, claude-devtools reconstructs a single Claude Code session down to which category of context ate each turn, and Agent Analytics answers a different question entirely, how are my product’s users behaving, with the same agent-readable stance.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified. Each column links to the full research note; every cell traces to a source cited there or in the references.

The matrix
#

Feature Agent Analytics agents-observe agentsview AgentTrace ccusage claude-devtools claude-replay ClawTrace CodeBurn ctx Memex Token Monitor Tokscale
Kind agent-readable web product analytics, tracker.js events served to agents through CLI, MCP, and API real-time observability dashboard for live and replayed sessions local session indexer, web UI + CLI + desktop local-first TUI/CLI audit of coding-agent session history (cost, tokens, latency, failures, health) local cost-report CLI over the session files coding agents already wrote local desktop DevTools that reconstruct Claude Code sessions from ~/.claude logs (per-turn token attribution, tool-call inspection, subagent trees) local transcript-to-replay compiler producing one self-contained interactive HTML artifact per session hosted trace and cost-attribution platform for OpenClaw runs, with an AI analyst local spend-analysis desktop app plus CLI with task and branch attribution, plan quota, and spend guards local session search CLI with built-in blame attribution local session search CLI and TUI with resume-in-place and an MCP server live desktop usage widget and dashboard for tokens and provider limits across 43-plus AI coding tools terminal CLI and TUI reporting tokens and cost across about 50 local agent data stores, with an opt-in public leaderboard
Deployment self-hosted OSS server (Cloudflare Workers + D1, Docker/Kubernetes + SQLite, plain Node) or hosted cloud at app.agentanalytics.sh Claude Code plugin, hooks feeding a Dockerized local API server, dashboard on localhost:4981 local daemon, Docker, optional PostgreSQL push for teams single Rust binary, TUI plus CLI reports, Homebrew/npm/curl/cargo installs (winget dropped in v0.10.0) zero-install CLI (npx/bunx, or npm global) with JSON config; no daemon, no server desktop app (macOS/Linux/Windows; Homebrew, direct download, Docker on localhost:3456), reads ~/.claude in place zero-dependency CLI (npm global), a fully client-side web converter, and a Docker mode with read-only session mounts OpenClaw plugin streaming to a hosted cloud pipeline (Azure Blob, Databricks, PuppyGraph, Vercel UI), no documented self-host path CLI (npx, npm, brew) plus a signed desktop app (macOS, Windows Store, Linux deb/rpm/AppImage) with menu bar, tray, and GNOME plugins; no daemon single-binary CLI install, agent-callable skill single-binary CLI install (brew, AUR, Nix, cargo-binstall), TUI, local web UI, native desktop apps, MCP server desktop app (Windows 10+/macOS 12+/Linux x64, Homebrew cask) plus iOS widgets through the Worker hub; single-device use needs no server zero-install CLI (bunx/npm), TUI, visualization dashboard, sync subcommands for API-backed sources (Antigravity, Trae, Warp)
Open source ~ README claims MIT, no LICENSE file on the default branch as of 2026-09-27, sibling repos are MIT ✓ MIT ✓ MIT ✓ MIT ✓ MIT (license file at apps/ccusage/LICENSE in the monorepo) ✓ MIT ✓ MIT ~ Apache-2.0 repository, hosted cloud is the product and no self-host path is documented ✓ MIT ✓ Apache-2.0, blame included since 2.0 ✓ MIT ✓ MIT ✓ MIT
Agents covered any agent that can run commands or call HTTP, with documented installs for Claude Code, Codex, Cursor, OpenClaw, Paperclip, Hermes, Instinct, and OpenWork ~ Claude Code and Codex, with the plugin install Claude Code-native 60+ formats auto-discovered (Claude Code, Codex, Gemini CLI, Copilot, Cursor, Zed, OpenCode, and more) ~ about 13 named formats plus generic JSON/JSONL (Claude Code, Codex, Qwen Code, OpenCode, OpenClaw, Cursor exports, and more); Gemini CLI removed in v0.10.1 as discontinued upstream 18 named sources in the unified report (Claude Code, Codex, OpenCode, Amp, Droid, Codebuff, Hermes Agent, pi, Goose, OpenClaw, Kilo, Kimi, Qwen, Copilot CLI, Gemini CLI, Antigravity, Grok Build CLI, ZCode) ~ Claude Code only (terminal, IDEs, anything writing ~/.claude) ~ 7 formats auto-detected (Claude Code, Cursor, Codex CLI, Gemini CLI, OpenCode via export, Kimi Code, Hermes Agent via SQLite or export) ~ OpenClaw only, through its eight-hook plugin ~ 41 integrations claimed (Claude Code, Codex, Cursor, Gemini, and more); Cursor coverage reads the IDE database only, so Cursor Agent CLI transcripts are missed ~ about 40 agent harnesses documented (Claude Code, Codex, Cursor, Pi, OpenCode, and more) ~ 17 engines graded per capability (Claude Code, Codex, Cursor, OpenCode, Pi, GitHub Copilot CLI, Grok, Hermes, KiloCode CLI, and more), support uneven ~ 43-plus tools graded per capability (Claude Code, Codex, OpenCode, Cursor, Antigravity, Copilot, and more); several limits sources are account-level cookie reads (Ollama Cloud, Alibaba Cloud), not session files ~ about 50 named local sources (OpenCode, Claude Code, OpenClaw, Codex, Copilot CLI, Gemini CLI, Cursor, Amp, Droid, Pi, Kimi, Qwen, Cline, Goose, Crush, Zed, Kiro, Devin, Junie, and more), several through cached account exports or estimation
Token cost reporting ✗ product events, not agent token cost ✓ per-session token usage and cost breakdowns (since v0.9.7) ✓ per-model pricing catalog, seconds over months of sessions ✓ per-session tokens and estimated USD cost across sources, with pricing overrides and confidence levels ✓ daily, weekly, monthly, and session reports plus Claude 5-hour blocks and a beta statusline, with per-model breakdowns; pricing-catalog estimates ✓ per-turn token attribution across 7 context categories with compaction visualization, plus per-subagent tokens, duration, and cost ✗ ✓ per-step tokens and USD cost, 80+ models with cache-aware pricing ✓ by tool, model, project, git branch, and task, with plan-quota tracking; API-equivalent estimates ✗ ~ per-engine token usage, off by default ✓ live tokens and cost in USD/TWD/HKD/CNY with daily exchange rates, cache hit versus miss breakdowns, CSV and JSON export, manual subscription cost records ✓ tokens and cost from LiteLLM’s pricing catalog with cache discounts, daily and per-model views
Search ~ flexible analytics queries (metrics, group_by, filters), funnels, paths, retention, not transcript search ✓ filtering and search across live and stored events ✓ FTS5 full text, semantic search opt-in ✓ sort and filter sessions by cost, duration, health, failures, anomalies, model, source, or text ✗ report tables and JSON output, no transcript search ✓ command-palette cross-session search, multi-pane side-by-side sessions ✗ chapter navigation and bookmarks inside a replay, no cross-session search ~ trace tree, call graph, timeline browse, and natural-language Ask Tracy queries over the trace graph, not transcript search ~ sessions drill down to turns and periods compare side by side, no full-text search ✓ cross-agent message and tool-call search, subagent and fork aware ✓ BM25 default, opt-in local or remote (OpenAI-compatible) embeddings for semantic and hybrid queries, saved memories searchable, SSH federation ~ per-session drilldowns, trend and heatmap dashboards, no full-text transcript search ✗ report tables and TUI views, no transcript search
Provenance ✗ ✗ ✗ ~ heuristic Git delivery correlation only, no line-level attribution ✗ ✗ session-level inspection only, no code-line or commit attribution ✗ ✗ ~ spend matched to pull requests and git branches, not line-level attribution ✓ blame maps a line, file, commit, or PR to the session that produced it ✗ ✗ ✗
Live observation ✓ real-time terminal dashboard across projects, plus opt-in web session replay ✓ hook events stream to the dashboard over websockets as agents run ✗ retrospective only, parses files already written ✗ retrospective only, reads logs already written ~ live Claude 5-hour block monitoring and a beta statusline, not a session dashboard ✗ retrospective, reads transcripts already on disk ~ --serve --watch follows a running session and renders the replay as it goes ✓ runs stream to the hosted dashboard through the eight-hook plugin ~ menu bar, tray, and Capacity Dock show current spend and what is running now, not agent actions ✗ retrospective only ✗ retrospective only ✓ widget updates within seconds of each turn, live tok/s rate, menu bar and tray popovers ✗ retrospective parsing
Team features ~ multi-agent access and cross-project portfolios, no team roles or SSO documented ✗ single-user local setup ✓ PostgreSQL push, machine-labeled sync, S3 roots, versioned exports ~ CI reports and shared baseline artifacts, no hosted team sync ✗ ✗ single-user local, SSH inspection of remote machines ✗ the artifact itself is the sharing model, no team product ~ multi-tenant accounts (Tenant to Agent to Trace to Span) with referrals, no team roles documented ✗ ~ beta history sharing through a self-hosted ctx server with collection permissions and revocable access since v2.2.0, no hosted team product ✗ single-user, SSH federation to your own machines, object-storage sync announced ~ real-time multi-device sync through a self-hosted hub (in-widget, Node CLI, or Cloudflare Worker) or iCloud Drive, no team roles ✗ single-user; the opt-in leaderboard publishes public usage profiles
Privacy posture self-host keeps events in your D1 or SQLite, hosted cloud stores them in the vendor database, replay is opt-in and PII-masked local-first, events stay in a local SQLite store behind a local Docker server local-first, one anonymous ping by default, disableable local-first, prompt and result bodies are not stored in tool steps, history of derived metrics is opt-in local reads only, no accounts; fetches the model-pricing catalog unless run with –offline local reads only, nothing leaves the machine; SSH reads are user-initiated local reads with automatic secret redaction before export; the web converter runs entirely client-side in the browser hosted, trace payloads including LLM inputs and outputs are uploaded to the vendor’s data lake local reads, no account; guard and optimize edit local config with backups, and the MCP server pseudonymizes project names local-first, attribution refuses data not on the machine local-first, no cloud service in the default path, embeddings local by default (the opt-in remote path sends transcript text to that API) local-first, per-session transcript reads happen on demand and are never synced, sync carries usage not transcripts local reads by default, tokscale submit uploads usage data to build a public leaderboard profile (opt-in)
Pricing free cloud tier (100k events/month, 2 projects), metered cloud at $1 per 10k events, self-host free free, MIT, no paid tiers published free, MIT, no accounts free, MIT, no paid tiers published free, MIT, sponsor-funded, no paid tier free, MIT, the website lists a $0 offer free, MIT, no paid tier consumption credits; 100 free, packages $10 to $400, storage 1.35 credits/MB/day free, MIT, no paid tiers published free, Apache-2.0, the former $20/month pro add-on withdrawn from the site free, MIT, no paid tier published free, MIT, no paid tiers published, self-hosted sync free, MIT, sponsor-funded, no paid tier

Reading the matrix
#

The live-observation row is the row this matrix existed to name, and it now has three coding-agent answers: agents-observe streams hook events into its dashboard while the agents run, ClawTrace fills the same row for OpenClaw by uploading the run, and Token Monitor renders 43-plus tools’ usage and limits in one always-on widget as turns happen; Agent Analytics offers a real-time terminal dashboard for its own product events, though the hook-and-stream coding-agent cells still cover only Claude Code, Codex, and OpenClaw, so the remaining empty cells belong to the harnesses none of them hook. The second row worth reading is provenance: ctx’s blame attribution is the only cell in the category that answers “which session wrote this”, which agentsview deliberately leaves to cost and history questions.

Token cost reporting is the row that pays for the tool: harness-native cost views reset and see only their own sessions, while a pre-indexed store answers multi-tool, multi-month questions in seconds. An earlier version of the agentsview docs benchmarked its reports at 84 to 223 times faster than ad-hoc parsing (calling that an upper bound); the current docs have dropped that benchmark entirely, so the row should be read as “fast because pre-indexed”, with no vendor number left to lean on. The docs’ token-usage page documents schema version 5 again as of 2026-10-07, after documenting version 6 on 2026-10-06, the fourth flip this section has recorded, so scripts consuming those reports should expect churn either way. AgentTrace adds a cost cell the archive tools do not, with explicit estimate labeling and pricing overrides, while ClawTrace adds per-step cost inside a hosted trace, which is more granular than anything local but only for OpenClaw. ccusage and CodeBurn make cost the whole product: ccusage is the zero-install incumbent the other tools benchmark against, CodeBurn adds task and branch cuts plus an optimize loop that proposes config fixes, and both print estimates computed from pricing catalogs rather than bills. Tokscale and Token Monitor extend the same estimate approach to about 50 and 43-plus sources respectively, Tokscale on LiteLLM’s catalog and Token Monitor adding currency choice and manual subscription cost records.

Memex adds the closing move the other archive tools lack: resume-in-place, where finding a session and re-entering it are one action in the TUI, plus the category’s only semantic search (BM25 by default, opt-in embeddings that run locally or, since 2026-10-03, against an OpenAI-compatible API that receives your transcript text, so the headline requires setup either way). Its 17-engine support table is graded per capability and unevenly at that (some engines have no resume at all, token counting is missing for at least one), and its launch footprint is as thin as ctx’s was, so read the column as promising and unproven.

The 2026-09-27 columns split the category’s remaining questions: AgentTrace is the local cost, latency, and failure auditor that adds slow-run diagnosis the archive tools skip, ClawTrace is the hosted answer that trades privacy for full LLM payloads and an AI analyst, and Agent Analytics tests where the category’s boundary sits by serving product analytics to agents instead of session data.

The 2026-10-06 columns push on three remaining fronts: Tokscale widens the ledger to about 50 declared sources and adds the category’s first opt-in public leaderboard, Token Monitor makes the widget live and multi-device and keeps history after harnesses prune their transcripts, and claude-devtools shows how deep one harness’s logs go when the terminal stops telling you, per-turn token attribution across seven context categories included.

claude-replay adds the output format the rest of the category never produces: a single dependency-free HTML file you can email, host, or embed, compiled from seven harnesses’ transcripts with secret redaction applied, which makes it the only member whose reader is someone who never installed the tool, and its ledger cells (cost, search, provenance) are empty by design rather than by gap.

Breadth of coverage is agentsview’s moat: roughly 60 supported sources against Tokscale’s about 50, Token Monitor’s 43-plus, CodeBurn’s claimed 41, ctx’s 40, ccusage’s 18, Memex’s 17, AgentTrace’s 13, claude-devtools’ single harness, and the two harnesses agents-observe covers or the one ClawTrace hooks, which matters because most practitioners now run two or three harnesses at once. Agent Analytics is agent-agnostic by design, since any agent that can call HTTP is a client. ctx’s moat is different: agent-facing retrieval, where the consumer of the search is your next session rather than you.

Choosing from the matrix
#

  • Running three or more different coding agents and wanting one private history and cost view: agentsview.
  • Wanting your agents to recall why code exists, from the session that wrote it: ctx, with its self-reported numbers accepted.
  • Living in the terminal and wanting semantic recall plus one action back into the session: Memex, accepting a few hundred stars and thin verification.
  • Auditing what a run cost and why it was slow, locally and with CI gates: AgentTrace.
  • Wanting the quickest terminal answer to what did my agents cost this week, without installing anything: ccusage.
  • Wanting spend cut by task and branch, plan-quota tracking, and a guard that stops a runaway session: CodeBurn.
  • Wanting the widest declared token ledger, and maybe a public profile of your usage: Tokscale, after reading its submit flow first.
  • Wanting one live desktop view of tokens and limits across every tool, synced across your machines: Token Monitor, accepting a fast churn rate.
  • Wanting to know why a Claude Code context window filled and exactly what a tool call did: claude-devtools.
  • Running OpenClaw and wanting full-payload traces plus an AI analyst, hosted: ClawTrace, accepting that the traces leave your machine.
  • Instrumenting a product your agents are building and reading its traffic through an agent: Agent Analytics.
  • Needing to show a session to someone who has no tooling installed: claude-replay, compiling the transcript into one HTML file they can open anywhere.
  • Needing to see what an agent is doing right now: agents-observe for Claude Code or Codex, or ClawTrace if you run OpenClaw and can upload the run; otherwise a harness-native view is still the fallback.
  • Single-agent users: your harness’s built-in usage views are probably enough.

Changes
#

  • 2026-08-30 - Created as the Session analytics category’s companion matrix, a single-column scaffold with the live-observation gap named in prose.
  • 2026-09-05 - Extended from one to two columns with ctx.
  • 2026-09-06 - Updated the ctx cell for its published pro pricing.
  • 2026-09-08 - Dropped the agentsview benchmark claim from the prose after the tool’s docs removed it entirely.
  • 2026-09-13 - Added the agentsview usage-output schema version 6 churn note to the reading prose.
  • 2026-09-16 - Extended to three columns with agents-observe, filling the live-observation gap the prose used to name.
  • 2026-09-18 - ctx agents-covered cell updated after the docs began listing about 40 supported agent harnesses.
  • 2026-09-20 - Extended from three to four columns with Memex, inserted last alphabetically, every row gaining a cell traced to the new note, with the reading and choosing prose extended.
  • 2026-09-20 - Corrected the schema caution in the reading prose: the docs’ usage JSON contract is at schema version 5, version 6 is the session-export schema.
  • 2026-09-24 - Removed the verification preamble line on owner request.
  • 2026-09-25 - Retired the ctx pro cells: the Kind, Open source, Team features, and Pricing cells now describe blame as built in, after v2.0.0 shipped it inside the single open-source executable and the site dropped every paid-tier listing.
  • 2026-09-27 - Re-verified all four columns: repository counts refreshed (agents-observe 684, agentsview 6,001, ctx 1,135, Memex 231 stars), the reading prose’s schema line revised after the agentsview docs returned to version 5, and the Memex star count in the choosing prose generalized; no table cells moved.
  • 2026-09-27 - Extended from four to seven columns with Agent Analytics, AgentTrace, and ClawTrace, re-sorted all columns alphabetically, updated the intro member count and framing, and extended the reading and choosing prose.
  • 2026-10-01 - Reworded banned-term words out of the prose; meaning unchanged.
  • 2026-10-02 - Moved the ctx Team features cell after v2.2.0 shipped opt-in history backup plus beta sharing through a self-hosted ctx server; the other six columns re-verified unchanged.
  • 2026-10-03 - Updated the Memex agents-covered cell and the prose engine counts to 17 engines after v0.25.0 (2026-10-02) added Hermes and KiloCode CLI; the other six columns re-verified unchanged.
  • 2026-10-04 - Updated the Memex Search and Privacy posture cells and the reading prose after v0.26.0 (2026-10-03) added opt-in OpenAI-compatible remote embeddings alongside the local path; the other six columns re-verified unchanged.
  • 2026-10-05 - Extended from seven to nine columns with ccusage (the zero-install cost-report incumbent, inserted in sorted position between AgentTrace and ClawTrace) and CodeBurn (the desktop spend-analysis app with task attribution and spend guards, between ClawTrace and ctx), every row gaining two cells traced to the new notes; the AgentTrace Deployment cell dropped winget after v0.10.0 removed it, the intro count and reading and choosing prose updated.
  • 2026-10-06 - Extended from nine to twelve columns with claude-devtools (the single-harness Claude Code session DevTools), Token Monitor (the live desktop usage widget with provider limits and self-hosted multi-device sync), and Tokscale (the about-50-source token ledger with an opt-in public leaderboard), inserted in sorted position, every row gaining three cells traced to the new notes; the AgentTrace agents-covered cell dropped Gemini CLI after v0.10.1 removed it, the intro count and the reading and choosing prose updated, and the agentsview schema note moved back to version 6 per the current docs.
  • 2026-10-07 - The reading prose’s schema line moved back to version 5 after the agentsview docs flipped again (version 6 documented on 2026-10-06, version 5 on 2026-10-07), the fourth recorded flip; no table cells changed.
  • 2026-10-07 - Extended from twelve to thirteen columns with claude-replay (the transcript-to-replay compiler producing self-contained HTML artifacts across seven harnesses), inserted in sorted position between claude-devtools and ClawTrace and traced to the new note; the intro reframed the category as four arrangements with the shareable artifact as the new one, and the reading and choosing prose gained the claude-replay entries.

See also
#

References
#