SolonGate Agent Security stops AI coding agents from reading .env files

A coding agent with shell access can read a password file in one tool call. The user has no way to see what left the machine.

SolonGate Agent Security is an open-source gateway. It sits between the agent and the shell. The README describes it as a local-first security gateway for AI coding agents. Each tool call is checked against your rules before it runs. It also looks for secrets, caps how often each call runs, and writes a local audit log. The project is in Go and licensed Apache-2.0. The repository was created on 27 September 2026 and now holds 256 stars. It ranks 8th on Trendshift's daily list on 7 November 2026.

What it checks

The gateway runs rules by file path, command type, filename pattern and URL. Each rule is scoped to READ, WRITE, EXECUTE or NETWORK. A call that fails a rule is refused. The agent sees the refusal. The README shows a Claude Code session. The agent asks for a .env file. The guard refuses. Claude reports the block and does not try another path. Each refusal and each allowed call is written to an on-device audit log. The agent is not given access to that log.

A second layer looks inside each allowed call for secrets. The README says the project carries "70+ built in secret patterns plus your own." Each pattern has four modes: off, detect, redact, or block. Redact mode lets a call go through with the secret replaced. The agent can keep working on a file that contains a key. The key does not leave the machine.

Rate limits run on the same calls. The gateway caps how many times per minute, hour or day each tool class can run. An agent that goes into a loop stops hitting an API before the bill grows.

Which agents it works with

The README lists four agents as supported today: Claude Code, Codex CLI, Antigravity CLI and OpenCode. Three more are marked as coming: OpenClaw, Hermes and Chainabit. Installation runs through ./scripts/install.sh. The script checks for Go 1.26 and installs the runtime if the user does not have it.

What the project does not claim

SolonGate does not claim to make an agent safe. It publishes no benchmark of blocked attacks. The 256 stars and the Trendshift ranking are the only outside signals of interest so far. A team that cares about a specific risk has to write the rule for it. The 70 built-in patterns catch common API keys. A project that stores credentials in an odd file still needs its own pattern.

Running the gateway also adds a step to every tool call. The README does not give a measured latency. A team weighing it against a more open setup should test the cost on its own workflow.

Source

• solongate/agent-security, SolonGate on GitHub.
• Trendshift daily ranking, 7 November 2026.