Mats2208/MCP-Packet-Tracer: AI-powered Cisco Packet Tracer automation over MCP. Describe a network in natural language and it generates, validates and deploys full topologies (OSPF, EIGRP, RIP, DHCP).

Describe a network in plain words. Your AI plans it, validates it, and builds it live in Cisco Packet Tracer — then pings it to prove it works.

_{Install it by asking your agent: "Install the Packet Tracer MCP from github.com/Mats2208/MCP-Packet-Tracer and set everything up for me."}

79 MCP tools

6 MCP resources

3 MCP prompts

74 device models

150 modules

15 cable types

_{▶ Watch on YouTube (6 min): install, how it works, and three live examples. Spanish narration, English subtitles.}

Showcase

_{From a prompt to a running topology in Packet Tracer, in real time}

How it works

"3 routers, OSPF and DHCP"
        │
        ▼
  Your AI agent ──── Claude Code, OpenCode, Codex, Cursor… (any MCP client)
        │  calls pt_* tools
        ▼
  Packet Tracer MCP ─ plans → validates → auto-fixes → generates JS + IOS config
        │  authenticated HTTP bridge, 127.0.0.1:54321
        ▼
  MCP Control Center ─ this project's PT extension, polls the bridge every 500 ms
        │  runs the JavaScript in PT's Script Engine
        ▼
  Cisco Packet Tracer ─ devices, cables and configs appear on the canvas

With the extension window closed, a file mailbox takes over, so PT keeps executing. Everything stays on your machine: the bridge only listens on loopback and every call carries a per-machine token (see Security).

What it can do

| | What | Tools | |---|---|---| | Build | One prompt → plan → validation → auto-fix → live deploy, with read-back of every device and link | pt_full_build, pt_plan_topology, pt_validate_plan, pt_fix_plan, pt_live_deploy | | Addressing | /24 LANs, /30 point-to-point links, DHCP pools on routers or on a dedicated Server-PT, IPv6 dual stack | pt_configure_dhcp_server, dual_stack=True | | Routing | Static (with floating backups), OSPF, EIGRP, RIP | generated IOS, applied live | | Switching | VLANs, trunks, router-on-a-stick, STP, port security | pt_apply_vlan, pt_apply_stp, pt_apply_port_security | | Security | ACLs, NAT, device hardening (SSH, local users, enable secret, banner) and an audit of the live config | pt_apply_acl, pt_apply_nat, pt_apply_hardening, pt_audit_security | | Inside each device | Type in a router's CLI tab, run ping/ipconfig on a PC, configure Server-PT DHCP/DNS/HTTP/FTP/EMAIL, use the Web Browser and Email apps | pt_cli, pt_host_command, pt_server_*, pt_web_browser | | Verify | Real ping with parsed results, health check, plan-vs-live diff, per-port inspection | pt_verify_connectivity, pt_health_check, pt_diff, pt_inspect_ports | | Explain | Step the simulation and read PT's own per-layer decision log for each packet | pt_simulation_mode, pt_simulation_step, pt_read_packet_trace | | Document | Canvas screenshots, notes and drawings on the canvas, .pkt save/open, config backup, exported plans and configs | pt_screenshot, pt_add_note, pt_save_project, pt_backup_config, pt_export |

Every tool is listed in the tool reference.

Installation

Needs Cisco Packet Tracer 8.2+ (tested on 9.0 and 9.0.1) on Windows or macOS — see Platforms.

The easy way: ask your AI agent

Open Claude Code, OpenCode, Codex or any coding agent and say:

Install the Packet Tracer MCP from https://github.com/Mats2208/MCP-Packet-Tracer and set everything up for me.

That's it. The agent finds its instructions here, checks your OS and agent, installs the server and the skill, and puts the extension where Packet Tracer finds it. The next time you open Packet Tracer it asks "detected a new script module… enable?" — click Yes. That click is the only thing you do by hand.

[!NOTE]

AI agent reading this? Don't follow the manual steps below. Follow

llms-install.md — it is written for you: OS detection, every client's registration

command, the skill, the extension, and how to verify with the user.

Manual install

1. Install uv — it brings its own Python, so you don't need to install Python or touch PATH.

_macOS · Linux:_ curl -LsSf https://astral.sh/uv/install.sh | sh _Windows (PowerShell):_ powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Connect your MCP client (Claude Code shown; other clients in the Installation docs)

claude mcp add --scope user --transport stdio packet-tracer -- uvx packet-tracer-mcp --stdio

_Windows PowerShell_ — quote the -- separator, or PowerShell swallows it and Claude aborts with error: unknown option '-m':

claude mcp add --scope user --transport stdio packet-tracer "--" uvx packet-tracer-mcp --stdio

Verify with claude mcp list (look for packet-tracer … ✓ Connected).

3. Install the Agent Skill — _required in practice: without it the AI guesses model, port and API names instead of discovering them_

uvx packet-tracer-mcp install-skill                         # Claude Code, OpenCode → ~/.claude/skills
uvx packet-tracer-mcp install-skill --dest ~/.agents/skills # Codex

Then restart your agent (or run /reload-skills in Claude Code). Details → Skill docs.

4. Install the live-deploy extension — Download MCP-Control-Center.pts from Releases and drop it in Packet Tracer's own extensions folder, in a subfolder (…/extensions/MCPControlCenter/). On macOS that is Cisco Packet Tracer .app/Contents/extensions/; on Windows C:\Program Files\Cisco Packet Tracer \extensions\ (needs admin). Next time PT starts it asks "detected a new script module… enable?" — click Yes. Or add it by hand with Extensions → Scripting → Configure PT Script Modules → Add…. Full walkthrough in Live deploy below.

v0.6.0+ requires V5. The bridge authenticates with a per-machine token that the V5 extension reads automatically; builds before V5 can't authenticate.

pip install packet-tracer-mcp            # then use python -m packet_tracer_mcp --stdio as the command
pip install "packet-tracer-mcp[ui]"      # Windows: adds UI mode (opening and capturing device windows)

Quick start

Restart your agent after installing, open Extensions → MCP BUILDER in Packet Tracer, and ask:

"Create a network with 1 switch and 2 PCs, deploy it to Packet Tracer and ping between the PCs."

"Build 3 routers in a triangle with OSPF, a LAN with 2 PCs behind each one, and DHCP."

"Check the network: show me R1's routing table and audit the security of every router."

The agent calls pt_full_build (plan → validate → generate → deploy), then verifies with a real ping. More in the Quick Start guide.

Platforms

| | Windows | macOS | Linux | |---|---|---|---| | Install by agent (llms-install.md, PT detects the extension) | ✅ | ✅ | not tested yet | | Server (uvx packet-tracer-mcp) | ✅ | ✅ | not tested yet | | Live deploy (HTTP bridge) | ✅ | ✅ | not tested yet | | File-bridge (extension window closed) | ✅ | ✅ | not tested yet | | Device panel (pt_cli, pt_host_command, services…) | ✅ | ✅ | not tested yet | | UI mode (open/capture device windows) | ✅ | — | — |

macOS was verified end to end on Packet Tracer 9.0.1: all 79 tools were run against a live canvas, and the bugs that run found are fixed in this release. Windows 11 was verified the same day on PT 9.0.1: the agent install (extension copied to Program Files, one UAC prompt, PT's "enable?" dialog), both bridge channels on the first call, OSPF + DHCP builds checked with real pings, and UI mode capturing device windows. Linux runs the same code path as macOS; if you try it, an issue with the result is very welcome.

Live deploy

Live deploy runs through this project's own Packet Tracer extension, the MCP Control Center (V5.2.pts in Releases). Add it once via Extensions → Scripting → Configure PT Script Modules → Add…, then open Extensions → MCP BUILDER in each session. It connects to the bridge by itself and reads the token from disk — there is nothing to paste or configure.

_{Adding the MCP Control Center extension (V5) to Packet Tracer}

📖 Full steps → Live Deploy Setup.

Device panel and UI mode

Once a topology is live, the agent works inside each device the way a student would: it types in a router's CLI tab, runs commands in a PC's Command Prompt, sets IP Configuration, browses with the Web Browser and configures a server's Services — all through Packet Tracer's own API, so what it types shows up in the real CLI tab and you can take over at any time.

It is headless by default. On Windows, say "show it in Packet Tracer" (or call pt_ui_mode("ui")) and every panel tool also opens the device's window on the matching tab; capture=True saves a PNG. Windows are driven with UI Automation and a click posted to the canvas — your real mouse and keyboard are never used.

Clients

Any MCP client drives this server — Claude Code, Cursor, Claude Desktop, VS Code with Copilot, Codex. Nothing in it is client-specific.

There is also one built on it: PacketSmith, a terminal app that runs these tools with the network drawn beside the conversation — a fabric tree and a canvas plan derived from the pt_* results themselves, so a device the model says it created but did not never shows up.

| | This server alone | This server + PacketSmith | |---|---|---| | Where you talk | the MCP client you already use | a terminal app built for this one job | | What you see | a chat log, plus PT in another window | split screen: reply left, live topology right | | Topology | read out of the tool output | fabric tree and canvas plan, drawn for you | | Tools loaded | every MCP server you have configured | this one only — measurably faster to start | | Engine | whatever your client speaks | Claude, or ~150 providers |

You need this server either way — PacketSmith runs it underneath, and its setup installs it for you. If you already live in Claude Code, you are done here; the client is for when you want the topology in front of you instead of buried in a scrollback.

npm i -g packetsmith    # beta · MIT

Credits & Acknowledgements

Live deploy runs through our own Packet Tracer extension — the **MCP Control Center** (the .pts in Releases). Its Script-Engine helper layer was inspired by PTBuilder by Kim Knight (@kimmknight), who pioneered driving Packet Tracer's Script Engine from JavaScript — thanks for the groundwork. 🙏

PTBuilder and Packet Tracer MCP are separate, independent projects. You install

our extension, not PTBuilder. Full

Credits & Attribution.

Security

Driving Packet Tracer from outside means running a local HTTP bridge whose whole job is to hand JavaScript to PT's Script Engine — code that executes with PT's own privileges, including disk access. That makes the bridge a genuine attack surface, not an implementation detail, and it is hardened accordingly.

The attack this design exists to stop. Binding to 127.0.0.1 is not a security control. A POST with Content-Type: text/plain is a CORS *simple request*: any web page open in your browser can send it to a loopback port without a preflight and without needing to read the response. An unauthenticated bridge therefore lets any website you visit — while Packet Tracer happens to be open — queue arbitrary code inside it. Injection never needed to read anything back, so same-origin policy alone never closed this.

What actually closes it is a secret the attacking page cannot guess:

| Control | Implementation | |---|---| | Token on every endpoint | Every route except /ping requires a shared token (?t= or X-PT-Token). Compared with hmac.compare_digest — constant time, no early-exit oracle. | | /ping leaks nothing | Deliberately unauthenticated so the server can tell who owns the port before trusting it — but it returns only a SHA-256 fingerprint of the token, never the token. | | Foreign-bridge detection | Before sending any payload, the server checks that /ping identity matches its own token fingerprint. If a stranger holds the port, it refuses to hand code to it instead of blindly trusting a 200. | | DNS-rebinding defense | The Host header is validated against 127.0.0.1 / localhost / [::1] + the real port. A rebound request arrives as Host: evil.com: and is rejected. | | Loopback bind | ThreadingHTTPServer(("127.0.0.1", port)) — never 0.0.0.0, so the bridge is not reachable from the LAN. | | Token at rest | secrets.token_urlsafe(32), created with O_EXCL (race-safe when two servers start at once) at mode 0o600, under %LOCALAPPDATA% on Windows — deliberately not roaming %APPDATA%, so a loopback secret never syncs to a file server. | | Body size cap | Oversized bodies are rejected with 413 and are not read into memory. | | Silent failures | Error responses carry no CORS headers, so a hostile page cannot even distinguish why it failed. | | Tamper visibility | Unauthorized attempts are counted and surfaced by pt_bridge_status, so a stale or rogue client is diagnosable instead of silent. |

Regression coverage lives in tests/test_bridge_security.py and tests/test_injection_regressions.py; the full suite runs offline with python -m pytest — no Packet Tracer required.

v0.6.0+ requires the V5 extension. Versions before v0.6.0 shipped an

unauthenticated bridge and are vulnerable to exactly the attack above.

Upgrade — there is no safe configuration of the old bridge.

Deliberate, documented behaviour: pt_send_raw executes arbitrary JavaScript inside Packet Tracer by design — it is the escape hatch for exploring the IPC API. It is reachable only by an MCP client you have already authorised, over the authenticated bridge. That is a capability, not a vulnerability.

Found a vulnerability? Report it privately via GitHub Security Advisories, not a public issue. SECURITY.md documents the full threat model.

What's new

Unreleased (v0.10.0) — install it by asking your agent: llms-install.md walks any coding agent through the whole setup on Windows, macOS or Linux. The skill ships inside the package (uvx packet-tracer-mcp install-skill), uvx packet-tracer-mcp runs the server with no Python install, and 17 device-panel tools let the agent work inside each device window.

v0.9.0 — on PyPI, and a hunt for silent false OKs: a topology split into islands that validation approved, a bridge returning the previous operation's result, a connectivity "OK" with 75% packet loss. None of them crashed; all of them lied. Older releases in the Changelog.

Contributing

See CONTRIBUTING.md. The test suite (945 tests) runs offline with uv run --extra test python -m pytest — no Packet Tracer needed.

License

Released under the MIT License — © 2026 Mateo (@Mats2208).