Repository navigation
Releases
v0.4.0 — network trigger rules
Phase-3 alerting lands: pounce now classifies each correlated connection against the trigger rules in TRIGGERS.md, replacing the earlier ✓/? verdict with a type · severity · confidence finding model. Still the zero-FDA MVP (Phases 1–3); still observe-only.
Highlights
- Trigger rules —
⚠alert /?to review /✓expected, led by a one-line verdict:#1capability mismatch — a local-only tool (built-in prior + learned baseline) egressed at all#2destination mismatch — a call connected somewhere it didn't declare (eTLD+1 compare; the declared host's own IPs count as legit)#3cloud-metadata / raw-IP properties#4novel destination — a new registrable domain for a(server, tool), soft during its first-session learning window
- Learned baseline — persisted per-
(server, tool)profile at~/.pounce/profile.json, order-independent, learned onwrapsession close. - Confidence = attribution × naming; low-confidence findings stay in the timeline rather than pushing.
- Dedupe — one finding per
(tool call, destination), shown as×N.
The headline exfil demo now reads at a glance:
⚠ 1 alert — destination mismatch
tools/call fetch {"url":"https://example.com"}
✓ tcp 93.184.216.34:443 declared example.com
⚠ tcp 1.1.1.1:443 declared example.com — connected to 1.1.1.1 (no DNS)
The feature (#5) was adversarially reviewed before merge; 9 issues from that pass were fixed with tests.
Install (macOS)
curl -sSL -o pounce https://github.com/tstefank/pounce/releases/download/v0.4.0/pounce-darwin-arm64
chmod +x pounce && xattr -d com.apple.quarantine pounce 2>/dev/null || true(Use pounce-darwin-amd64 on Intel.) Verify with checksums.txt.
Full changelog: v0.3.1...v0.4.0
v0.3.1 — verdict-first view + parallel correctness
A UX + correctness release on top of the MVP. Same capabilities (Phases 1–3, zero FDA), but the output now reads at a glance and parallel servers are handled correctly.
Legibility
pounce view defaults to a verdict-first summary — a one-line verdict, then each tool call with the connections it caused:
node fetch-server.mjs <session>
? 1 connection to an IP with no DNS lookup
tools/call fetch {"url":"https://api.example.com"}
✓ tcp 93.184.216.34:443 api.example.com
? tcp 45.83.12.9:443 no DNS lookup — unverified destination
1 tool call · 2 connections · 1 unverified (--timeline for the full log)
- Two-level confidence:
✓confirmed (the connection went where the call declared),?unverified (a host the call didn't declare, or no DNS lookup at all). There's deliberately no red "⚠ malicious" verdict — passive correlation can't assert that (a no-DNS connection may just be a cached lookup), so pounce reports honest confidence and leaves hard alarms to a future rule layer. --timelinekeeps the full chronological protocol + OS log.--all [--limit N]rolls several sessions into one overview — the view when multiple servers run at once.
Parallel correctness
- Destination-matched attribution: a connection now attributes to the call whose arguments declared its host (via DNS), so concurrent tool calls to different hosts split correctly instead of both landing on the latest call.
- Unique session ids: two shims starting in the same millisecond no longer collide into one log (the id carries the PID).
- DNS is shared: resolutions are broadcast to every session (connections stay attributed by PID subtree), so a legitimate connection isn't mis-flagged just because
mDNSResponderdid the lookup.
Validated live: two servers wrapped at once (one trojan, one honest) show one ? and one ✓, correctly separated. scripts/demo-parallel.sh reproduces it.
See docs/correlation.md for the model and caveats (DNS-cache false positives, conservative declaration).
Install
Static binaries, no dependencies:
chmod +x pounce-darwin-arm64 # or -amd64 on Intel
xattr -d com.apple.quarantine pounce-darwin-arm64
mv pounce-darwin-arm64 /usr/local/bin/pounceVerify with shasum -a 256 -c checksums.txt.
Full changelog: v0.3.0...v0.3.1
v0.3.0 — MVP: intent↔effect correlation with divergence flagging
The MVP. pounce now closes the loop: declared intent (tool calls) → actual OS effect (network connections) → divergence. Other tools see one half — the tool call or the syscall. pounce ties them together and flags where they disagree. All of this is zero Full Disk Access.
14:00:00.000 -> tools/call fetch {"url":"https://api.example.com/data"} -> ok
14:00:00.050 · dns api.example.com → 93.184.216.34 node pid 999
14:00:00.200 · net tcp 93.184.216.34:443 out node pid 999
14:00:00.400 · net tcp 45.83.12.9:443 out node pid 999
correlation (window 3s):
tools/call fetch → caused 2
tcp 93.184.216.34:443 (api.example.com, declared) ← declared host → DNS → connection ✓
tcp 45.83.12.9:443 (unresolved — no DNS) ← hardcoded IP
⚠ 1 undeclared destination — IP not resolved from any host:
tcp 45.83.12.9:443 (unresolved — no DNS)
What's new (Phase 3 — correlation)
internal/correlatejoins each request to the connections that occurred during its execution window, and flags two kinds of divergence:- Out-of-band (timing): a connection while no tool call was active — the server reached out on its own.
- Undeclared destination (DNS): a connection to an IP that no DNS lookup produced — a hardcoded/exfil destination, caught even inside a benign call's window. This is the teeth: time-based attribution alone can't catch a rogue server that connects to a malicious IP under cover of a legitimate
fetch.
- The pktap parser now decodes DNS (query→response, joined by transaction id) into
host → IPs, so connections can be checked against what the server actually resolved. Declared hosts come fromhttp(s)://URLs in tool-call args (conservatively — we never over-declare and suppress an alarm). viewshows DNS resolutions and a correlation section;--windowtunes the attribution window.- Validated live on macOS: a real
example.comconnection ties back to its DNS lookup; a hardcoded1.1.1.1connection is flagged.
See docs/correlation.md for the mechanics and the honest caveats (DNS-cache false positives, DoH, correlation ≠ proof).
The full picture (Phases 1–3)
| Tier | Needs | Gives |
|---|---|---|
| Protocol timeline | nothing | declared tool calls |
Network capture (sudo pounce daemon) |
root, no FDA | actual per-process connections |
Correlation (pounce view) |
nothing | the join + divergence flags |
Install
Static binaries, no dependencies:
chmod +x pounce-darwin-arm64 # or -amd64 on Intel
xattr -d com.apple.quarantine pounce-darwin-arm64
mv pounce-darwin-arm64 /usr/local/bin/pounceVerify with shasum -a 256 -c checksums.txt. Build from source: go build -o pounce ./cmd/pounce.
Next
Phase 4: eslogger file/exec capture (root + FDA) — file reads and process spawns join through the same correlator. A rules/actions sink can then fire on a detected divergence.
Full changelog: v0.2.0...v0.3.0
v0.2.0 — OS ground truth: per-process network capture
The first release with OS ground truth. Beyond the declared tool calls (Phase 1), pounce can now show the actual per-process network activity those calls cause — attributed to the wrapped server's PID subtree.
sudo pounce daemon # opt-in privileged capture: root, NO Full Disk Access
pounce wrap -- <server> # auto-reports its subtree to the daemon; no privilege
pounce view # network events interleaved into the timeline13:01:28.517 · net tcp 104.20.23.154:443 out curl pid 50592 ← the connection a tool caused
What's new (Phase 2 — network tier)
pounceddaemon (sudo pounce daemon): runs the systemtcpdumpon thepktappseudo-interface as root, attributes every connection to a session's PID subtree, and streams it back over a local Unix socket. One capture stream serves all wrap sessions. Root-only, no FDA.- Privilege split per the design: the client-launched
wrapshim stays unprivileged and just reports its child PID subtree; the socket verifies peer credentials (LOCAL_PEERCRED). No daemon running →wrapbehaves exactly as in v0.1.x. viewinterleaves network events into the timeline and records capture provenance (tcpdump/macOS versions).- Hardened attribution: anchored parsing, IPv6, and fail-safe-to-"attribution unavailable" with a drift counter + a live CI self-check, so a future tcpdump/macOS format change trips CI, not users.
- Also includes the v0.1.x line: tool-execution-error surfacing, and a wrap stdin-hang fix.
A note on the capture mechanism (macOS 26)
Network attribution needs a per-packet PID. On macOS 26 / tcpdump Apple-158 the binary paths don't expose it: the public libpcap API cannot enable DLT_PKTAP (confirmed with a pure-C probe — not a Go bug), and written pcap/pcapng drops the pid. So pounce reads the pid from the system tcpdump's text pktap metadata (-k NP) — same kernel pid — with strict, drift-guarded parsing. Details and the future raw-BPF option are recorded in the repo.
Install
Static binaries, no dependencies. Download for your Mac, then:
chmod +x pounce-darwin-arm64 # or -amd64 on Intel
xattr -d com.apple.quarantine pounce-darwin-arm64 # if Gatekeeper blocks it
mv pounce-darwin-arm64 /usr/local/bin/pounceVerify with shasum -a 256 -c checksums.txt. Or build from source: go build -o pounce ./cmd/pounce.
Not included
File/exec capture via eslogger (Phase 4: root + FDA) and intent↔effect correlation (Phase 3) — next on the roadmap. Linux is later.
Full changelog: v0.1.1...v0.2.0
v0.1.1 — surface tool-execution errors
Patch release. Phase 1 (stdio shim + protocol tee), now hardened against real-world MCP traffic.
Fixed
viewnow surfaces tool-execution errors. MCP reports a failedtools/calltwo ways: a JSON-RPCerrorobject, or a successful response whoseresult.isErroristrue(the tool ran but failed). Previously only the first was flagged, so execution-level failures rendered as-> ok. They now show-> tool error: <message>and count in the summary.
Validated against real servers
Tested end-to-end through pounce against @modelcontextprotocol/server-everything (notifications, streaming progress, high-volume logging, resources, prompts), server-memory, server-sequential-thinking, and OpenClaw — both as an MCP client (OpenClaw → pounce → server) and as an MCP server (pounce wrap -- openclaw mcp serve). Forwarding remained byte-for-byte identical throughout.
Install
Download the binary for your Mac below, then:
chmod +x pounce-darwin-arm64 # or -amd64 on Intel
xattr -d com.apple.quarantine pounce-darwin-arm64 # if Gatekeeper blocks it
mv pounce-darwin-arm64 /usr/local/bin/pounceVerify with shasum -a 256 -c checksums.txt. Or build from source: go build -o pounce ./cmd/pounce.
Full changelog: v0.1.0...v0.1.1
v0.1.0 — Phase 1: stdio shim + protocol tee
First release of pounce — ground truth for what AI coding agents and MCP servers actually do.
Wrap any MCP server in one line and see every JSON-RPC tool call it makes. Observe-only: the bytes between client and server are forwarded byte-for-byte, so the wrapped server behaves exactly as if pounce weren't there.
pounce wrap -- npx -y @modelcontextprotocol/server-filesystem /tmp
pounce viewWhat's in this release (Phase 1)
wrap— transparent stdio tee: launches the real MCP server, forwards stdin/stdout/stderr unchanged, and records every JSON-RPC message to~/.pounce/sessions/<id>.jsonl.view— replays the session as a tool-call timeline: pairs each request with its response (direction-aware id matching), summarizesinitialize/tools/list/roots/listresults, with optional ANSI color.- Newline-delimited JSON-RPC 2.0 parsing confirmed against the MCP 2025-06-18 spec; handles legacy batch frames.
- Architected so an HTTP/SSE proxy and a log/hook ingester can slot in behind the same intent-source interface later.
No privileges required. macOS-first.
Install
Download the binary for your Mac below, then:
chmod +x pounce-darwin-arm64 # or -amd64 on Intel
xattr -d com.apple.quarantine pounce-darwin-arm64 # if Gatekeeper blocks it
mv pounce-darwin-arm64 /usr/local/bin/pounceVerify with shasum -a 256 -c checksums.txt. Or build from source: go build -o pounce ./cmd/pounce.
Binaries are unsigned (Phase 5 will add notarized/Homebrew packaging). Not yet on a Homebrew tap.
Not in this release
OS-level capture (files/network/processes) and intent↔effect correlation — see the roadmap in the README. Linux support is later too; these are darwin/amd64 + darwin/arm64 only.