Every tool call your AI agent makes — logged, timed, diff'd. Nothing sent anywhere.
AI coding agents read files, write code, run commands, and commit to git on your behalf. vigilo gives you a complete, queryable record of everything they do — without changing how they work or what they can do. Works with Claude Code and Cursor.
░ CLAUDE ░ df66fc59 16 Feb 17:26
│ my-app · feature/auth@23dbb5e
│ claude-opus-4-6
│ 17:26:39 ● EXEC Bash cargo test && cargo ins… 6.3s
│ 17:26:45 ○ READ Read src/auth.rs
│ 17:27:01 ○ READ Grep src/app/**/*.ts
│ 17:27:02 ◆ WRITE Edit auth.rs +12 -3
│ 17:27:03 ✖ ERR Bash npm install 2.1s
│ ··· 42 more calls ···
│ 17:35:01 ○ READ Read auth.rs
│ 17:35:06 ◆ WRITE Edit auth.rs +3 -1
│ 17:35:12 ● EXEC Bash cargo test 4.2s
└─ 52 calls · r:22 w:11 e:19 · 1 err · 48.3s
tokens: 65K in · 270 out · cache: 162K read · ~$0.03 (list pricing)
░ CURSOR ░ a3b8c91e 16 Feb 18:02
│ Auto
│ 18:02:14 ○ READ Read navbar.html 141ms
│ 18:02:19 ◆ WRITE Edit navbar.html +8 -3
└─ 2 calls · r:1 w:1 e:0 · 0.3s
tokens: 71K in · 5K out · cache: 905K read · $0.35 (2 reqs)
── vigilo stats ────────────────────────────────
76 sessions · 1215 calls · 1 error (0%) · 407s total
risk: 487 read · 329 write · 328 exec
tools files
───── ─────
303× Bash 84× hook.rs
270× Read 64× data.service.ts
240× Edit 62× view.rs
86× Write 30× main.rs
models
──────
310× claude-opus-4-6 996 in · 33K out · ~$66.07
168× Auto
144× claude-sonnet-4-5 102 in · 23K out · ~$3.09
total: 1K in · 56K out · ~$69.16 (list pricing)
░ CURSOR ░ you@example.com (pro_plus)
billing: 02-05 → 03-05 (user)
plan: 7000/7000 requests (0 remaining) 26%
── token usage (30d) ────────────────────────────
213 requests
1.2M input · 257K output · 38.9M cache read · 2.5M cache write
$18.89 total cost
by model
────────
153× Auto 921K in · 174K out · cache:28.5M · $10.41
40× claude-4.5-sonnet 176K in · 56K out · cache:4.7M · $6.26
20× composer-1.5 135K in · 26K out · cache:5.6M · $2.21
Launch a real-time web dashboard to visualize all agent activity.
vigilo dashboard # default port 7847
vigilo dashboard --port 9000 # custom port- Session timeline — calls, cost, and errors over time with interactive charts
- Live event feed — every tool call appears in real-time via SSE, with filtering by server/risk/tool and sortable columns
- Session merging — context-compressed conversation fragments are auto-merged into unified sessions
- Token breakdown — input, output, cache read/write by model with cost estimates
- Cross-dimensional stats — tools, files, models, and projects in one view
- Binds to
127.0.0.1only — never exposed to the network. Includes CORS, CSP, and host validation headers. If the port is in use, vigilo prompts for a random available port.
vigilo runs as a local MCP server over stdio. There is no network listener, no open port, no daemon. The AI agent (Claude Code or Cursor) spawns vigilo as a child process — the same way it would spawn any MCP server.
When the agent makes a tool call through MCP, vigilo logs the event to ~/.vigilo/events.jsonl (tool name, arguments, result, timing, risk level, git context, diffs) and then executes it locally. The agent sees no difference.
Some tools are built into the agent and bypass MCP entirely (Claude Code's Read, Write, Bash, Edit, etc.). For these, a lightweight PostToolUse hook pipes the event to vigilo hook via stdin after the tool completes. Same logging, zero overhead on execution.
Nothing leaves your machine. No telemetry, no phoning home, no accounts, no SaaS. The only exception is vigilo cursor-usage, which makes opt-in HTTPS requests to cursor.com to fetch your token usage — using your existing local Cursor credentials.
Encryption at rest is automatic. On first run, vigilo generates an AES-256-GCM key at ~/.vigilo/encryption.key and encrypts content (arguments, results) transparently. The key is zeroized from memory on drop. Metadata (tool name, risk level, timing, git context) is always plaintext so you can always see the shape of what happened. You can also provide your own key via VIGILO_ENCRYPTION_KEY env var.
curl -fsSL https://raw.githubusercontent.com/Idan3011/vigilo/main/install.sh | bashOr from source: cargo install --git https://github.com/Idan3011/vigilo.git
Then run the interactive setup:
vigilo setupAuto-detects Claude Code and Cursor, configures MCP servers and hooks, optionally generates an encryption key, and saves everything to ~/.vigilo/config.
For manual setup, see docs/manual-setup.md.
vigilo # MCP server mode (default, reads stdio)
vigilo dashboard # web dashboard on port 7847
vigilo view # ledger grouped by session
vigilo sessions # one-line session list
vigilo stats # aggregate stats across all sessions
vigilo errors # errors grouped by tool
vigilo tail # last 20 events (flat, chronological)
vigilo diff --last 1 # file diffs from last session
vigilo watch # live tail of incoming events
vigilo cursor-usage # Cursor token usage (last 30 days)
vigilo export # dump all events as CSV
vigilo summary # today at a glance
vigilo doctor # check configuration health
vigilo setup # interactive setup wizard
vigilo prune # delete old rotated ledger files
vigilo generate-key # generate AES-256 encryption keyMost commands accept --since, --until, --last, --session, --risk, --tool, and --expand flags. See docs/commands.md for full details.
- Command reference — all subcommands, flags, and examples
- Configuration — config file, environment variables, encryption
- Ledger format — JSONL schema, field reference, rotation
- Architecture — source tree, design principles, development
- Manual setup — Claude Code and Cursor JSON snippets