AI Traffic Control
Open-source guardrail plugin that scans every prompt, tool call, response and file read in coding-agent sessions against local rule packs before they take effect.
At a Glance
About AI Traffic Control
AI Traffic Control (ai-tc) is an open-source project from AKA Security that places one control point in front of events in an agent session. It scans prompts, tool calls, responses and file reads against installed rule packs on the user's machine, and installs as a plugin for Claude Code and Claude Desktop. The repository is licensed under Apache 2.0.
What It Is
ai-tc is a guardrail layer for agent harnesses. Each event is matched against rules, every match becomes a finding with a rule id, category, severity and matched span, and policy decides what happens next. Detection is mostly regex, with some keyword rules and validators such as a Luhn checksum for card numbers or a Shannon-entropy check for high-entropy secrets, which reduce false positives.
Built-in detections cover secrets (vendor keys, private keys, connection strings), PCI data (card numbers, IBANs, routing numbers), PHI and PII (medical record numbers, ICD-10 codes, emails, SSNs, phone numbers), and risky commands such as curl piped into a shell, force pushes, rm -rf on home or shared paths and credential file reads.
Policy Outcomes
A finding resolves to one of five outcomes: monitor (logged only, where most rules ship), warn (a warning in the session while content still goes through), redact (the matched value is replaced before it reaches the model, for tool inputs and outputs), block (the prompt or tool call is stopped with an explanatory message) and exception (one exact value let through, granted by hand from the aka CLI or dashboard). Detections can be promoted from monitor to warn, redact or block per rule or per category from the dashboard.
Install and Data Handling
In Claude Code, the plugin installs from the akasecurity marketplace; Claude Desktop adds the same marketplace through Settings. The standalone aka CLI provides the dashboard, local store and aka scan, and installs through Homebrew, Scoop, a bootstrap script or release archives.
Events are logged to a local SQLite file at ~/.aka/data/aka.db. Detection and enforcement run locally. Network paths are limited: an update notice, package-manager installs, the opt-in /aka:setup calibration that sends findings to the model API, and the opt-in aka attach, which registers a machine against a control plane the user's organization runs. The README notes that on a default install no detection ships on redact or block, so the local store can contain raw values until policies are promoted.
Community Discussions
Be the first to start a conversation about AI Traffic Control
Share your experience with AI Traffic Control, ask questions, or help others learn from your insights.
Pricing
Open Source
Apache-2.0 licensed core ai-tc guardrails, run locally on your machine.
- Apache License 2.0
- Scans prompts, tool calls, responses and file reads against rule packs locally
- Monitor, warn, redact, block and exception actions
- Local SQLite log and aka CLI with dashboard
- Installs via Claude Code plugin marketplace or Homebrew
Capabilities
Key Features
- Scans prompts, tool calls, responses and file reads before they take effect
- Detects secrets, PCI data, PHI/PII and risky shell commands
- Five policy outcomes: monitor, warn, redact, block, exception
- Per-rule and per-category policy promotion from a dashboard
- Rule packs defined as JSON rules with fixtures
- Validators such as Luhn checksum and Shannon entropy to reduce false positives
- Local SQLite store of events and findings
- aka CLI with dashboard, local store and aka scan
- Manually granted exact-value exceptions
- Optional attach to an organization-run control plane
