The only way to audit AI agents you didn't build.

OS-level attestation for third-party AI agents.
No SDK. No proxy. No cooperation required.

The visibility gap

When your developers use Claude Code, Cursor, or Copilot, you have no idea what these agents actually do.

What observability tools see

  • LLM API requests
  • Token usage
  • Response latency
  • Error rates

What you're missing

  • Files read and written
  • Commands executed
  • Network connections
  • Sensitive data access

How it works

Valsa uses eBPF to hook into the Linux kernel. It sees every syscall—regardless of what application made it.

$ sudo valsad --config /etc/valsa/config.yaml

[14:23:01] Valsa v0.1.0 starting
[14:23:01] Loading eBPF programs... done
[14:23:01] Attached: execve, openat, SSL_read, SSL_write
[14:23:01] Policy loaded: default (12 rules)
[14:23:01] Monitoring active

[14:23:15] claude-code (pid:4521) session started
  ├─ READ   ~/.ssh/id_rsa                    → DENY   [sensitive_file]
  ├─ EXEC   git status                       → ALLOW  [allowed_binary]
  ├─ EXEC   git diff HEAD~1                  → ALLOW  [allowed_binary]
  ├─ POST   api.anthropic.com/v1/messages    → ALLOW  [allowed_host]
  ├─ READ   src/main.rs                      → ALLOW
  ├─ WRITE  src/main.rs                      → AUDIT  [source_modification]
  └─ EXEC   curl https://pastebin.com        → DENY   [blocked_binary]

[14:23:15] Session attestation: sha256:a3f2c8...

Capabilities

Process Execution

Track every binary spawned by AI agents. See arguments, working directory, parent process.

File Access

Monitor reads and writes. Block access to sensitive paths like .ssh, .aws, .env.

HTTP Interception

See LLM API calls by intercepting TLS at the SSL library. No MITM proxy needed.

Rego Policies

Define allow/deny/audit rules. Hot reload without restart. Git-based distribution.

Cryptographic Attestation

SHA-256 chained audit log. Tamper-evident. Compliance-ready for EU AI Act, SOC 2.

Apache 2.0

Self-host everything. Audit the code. No vendor lock-in. Community-driven.

Get started in 60 seconds

# Install
$ curl -sSL https://valsa.ai/install.sh | sh

# Initialize in your project
$ valsa init

# Start monitoring
$ sudo valsad --foreground