Power your agent
คำสั่งของผู้เผยแพร่เอง ไม่ใช่เอนด์พอยต์ของ PlugMCP
Host a copy with our keys and logs.
npx -y @agentistics/[email protected]Claude Code ไม่มีลิงก์ติดตั้งสำหรับคำสั่งแบบโลคัล ให้คัดลอกคำสั่งด้านล่างแล้วรันในเทอร์มินัล
claude mcp add agentistics-mcp -- npx -y @agentistics/mcp@2.213.0รันในเทอร์มินัล
{ "mcpServers": { "agentistics-mcp": { "command": "npx", "args": [ "-y", "@agentistics/[email protected]" ] } }}claude_desktop_config.json
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
เปิดจากผู้เผยแพร่
agentistics reads what your AI coding assistants already write to disk and turns it into tokens, costs, sessions and activity you can actually reason about. Nothing is uploaded: on a single machine it never opens an outbound connection at all, and in Team Mode a machine sends computed metrics only — never chat.
It tracks six harnesses side by side:
Agentistics MCP server — Claude Code analytics for Claude Desktop and Claude Code
A metric a harness genuinely cannot produce renders as N/A, never as a confident 0 —
HARNESS_CAPABILITIES in @agentistics/core is the single source of truth for which is which.
curl -fsSL https://agentop.openvibes.tech/cli | bash
Or via npm — same binary, downloaded by postinstall from the same GitHub Release:
npm i -g @agentistics/agentop
Both install the identical Linux x86_64 binary; use whichever fits your workflow (npm is handy
when agentop is a dependency of another project's scripts).
Then just run it:
agentop
Bare agentop on a terminal opens the control center — one full-screen application in the
terminal's alternate buffer, so it adds nothing to your scrollback.
Its tabs are Services (start / stop / restart this machine, a central, or the Docker machine;
connect to or leave a central; enable a boot service), Setup, Logs, Commands, Help
and Contribute. Screens change with ←/→, panes with tab, and the footer always names the
keys that work in the current focus.
Open the dashboard at http://localhost:47292 (the API + MCP endpoint stays on 47291).
Download the latest .msi or .exe from the Releases page.
On first launch agentistics detects your Claude Code data path automatically (Windows native or WSL).
SmartScreen warning? "More info → Run anyway". The binary is not code-signed yet.
git clone https://github.com/agentistics/agentistics.git
cd agentistics && bun install
bun run dev # API (47291) + UI dev server (47292)
agentop opens the control center, and its Dashboard tab is a live dashboard that needs no
browser — Overview, Projects, History, Costs, Harnesses and Hardware, switched with the digits or
tab. (agentop tui is an alias for agentop.)Metrics group by normalized git remote (host/org/repo), independent of where the repo is
checked out or which machine produced the work. A repo's detail page carries Overview, Members,
Sessions, an Actions tab when it has CI runs, and a Dynamic Workflows tab that renders each
multi-agent Workflow run as a phase-by-phase timeline.
A tag is a saved grouping of sources — repositories, projects, machines, teams or accounts — optionally pinned to a date window. Tag responses are aggregate-only: counts and sums, never session rows or transcripts, with any key the viewer cannot see collapsed into an "other" bucket.
One machine runs as a central and aggregates usage from many members. Members push computed metrics only — raw chat is never stored centrally (it is fetched on demand over a reverse WebSocket, from the machine that holds it).
Every machine has a role: solo (local only, the default), central (the aggregator, a Docker service on port 48080), or member. A machine may be a member of several centrals at once, with different sharing rules for each.
# host a central
agentop central init && agentop central up
# join one — a token minted by a central with a public URL carries it,
# so the token alone is enough (--endpoint only for a bare token)
agentop member connect --token <token>
A central is not a shared password. First boot prints a one-time setup token and asks you to create the owner account; everyone else is invited. Accounts carry argon2id passwords, optional TOTP two-factor, and a role. A team is a scope key, not a label — being in a team is seeing it. Step-up authentication gates escalation, not paperwork: editing an account, deleting a team and changing a password ask for a second proof. Enrolling a machine, a token or a repository does not — a prompt people meet daily is a prompt they clear without reading.
Each connection carries its own rules across two dimensions (repositories and projects) in one of two modes: denylist (share everything except these) or allowlist (share only these). The rules never travel: the central is told nothing about them beyond a per-dimension count.
When several of your machines share one account, they tell each other what they withhold over a sealed machine-to-machine channel (X25519 → HKDF → AES-256-GCM, sender and recipient bound into the authenticated header, keys pinned on first sight and every new pin announced). That is what makes two things possible:
agentop ci-push, authenticated
by keyless GitHub OIDC against a registered-repos allowlist. The central stamps the repository
itself, so a runner cannot mis-report which repo it ran for. See docs/github-actions.md.A central can be published behind a tunnel. AGENTISTICS_EXPOSURE selects a profile —
local | lan | public — and public permanently revokes every route that touches the host
(shell, local chat, raw transcripts, MCP admin) and requires a second factor of every owner. The
profile is the only thing that decides a capability; no opt-in re-enables host power on public.
agentop doctor --exposed # run this BEFORE opening a tunnel
A check that could not be verified reports fail, never a reassuring pass. On a Docker central
run it from inside the container instead — ./central.sh doctor --exposed — where central.env is
the live environment and the database is reachable, so the owner-MFA and machine-token checks
actually run. On the host they cannot, and unverified counts as a failure.
agentopagentop hooks install installs two things into Claude Code — a skill that teaches it to split
independent work across several assistants and start them with agentop session batch, and a
SessionStart hook that tells each new session which of those are already running. A hook infers
nothing (it is a shell command on an event); the inference is Claude's, reading what the skill
teaches. Nothing is written to ~/.claude until you run that command, and uninstall takes back
exactly what it wrote.
→ Full reference: docs/cli.md · docs/claude-integration.md
Costs are computed per session from a pricing table with three layered sources, merged in order of trust: the built-in table (compiled in — the floor), the LiteLLM community dataset, and the vendors' own pages. A source that fails or returns junk costs freshness, never the ability to price anything; community rows that imply a unit change or sit more than tenfold from the built-in figure are dropped.
Settings → Pricing lists every model this machine has actually used, each with its origin
(official / community / builtin) — so a model nobody can price is visible rather than silently
guessed at.
Claude Code deletes session transcripts older than cleanupPeriodDays (30 by default) on every
startup. agentistics asks once, on first run, what to do about that:
Revived sessions appear in lists and agent metrics but never inflate aggregate totals.
Nay is a chat panel that answers questions about your own data by calling the MCP tools.
It runs claude --print under the hood, so it spends your Claude subscription quota.
The MCP server exposes the same analytics as structured tools (summary, harnesses, projects, sessions, costs, layout building, PDF export) and registers itself at user scope on first start — so any Claude Code session can use them, not just Nay.
claude mcp list # should show "agentistics"
/proc, a container that cannot see the host) it says so
instead of rendering an honest-looking zeroThe GIFs above are recorded from a demo fleet, never from a real machine — a published recording cannot be taken back, and a real project name or repository in one is a leak.
# 1. a pseudonymized fleet, derived from this machine's computed metrics
# (never from transcripts, so chat cannot reach it by construction)
bun run packages/server/scripts/seed-demo.ts --split 3 --force
# 2. run them
HOME=~/.agentistics-demo-home-1 PORT=47391 ./release/agentop server &
# 3. record
packages/server/scripts/record-all.sh # terminal → .cast + .gif
bun run packages/server/scripts/record-web.ts # web → .gif
Terminal recordings are asciinema casts (casts/*.cast, embeddable as a
real player with selectable text) rendered to GIF with agg.
See CONTRIBUTING.md. Everything in this project is in English — code, comments, commits, docs. Contributions are accepted under the Contributor License Agreement.
Agentistics is source-available under the Functional Source License 1.1 (FSL-1.1-ALv2),
and each release becomes Apache-2.0 two years after it ships. Use it personally or inside your
company, self-host it, modify it and build integrations for free. What the licence does not allow is
offering it to others as a competing commercial product or service: a hosted service, a rebrand or a
resale. Versions up to and including mit-final remain MIT. See LICENSING.md for the
full picture and TRADEMARKS.md for the name and logo.
Agentistics MCP server — Claude Code analytics for Claude Desktop and Claude Code
{ "mcpServers": { "agentistics-mcp": { "command": "npx", "args": [ "-y", "@agentistics/[email protected]" ] } }}เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
@agentistics/mcp ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-11-14 Agentistics MCP server — Claude Code analytics for Claude Desktop and Claude Code