AI clients and agents

Argent for Claude Code

Every Claude Code call and tool action runs through Argent: a named, owned agent, redaction, spend caps and a signed record. Here is what Argent governs, how it connects and what your team sets.

Anthropic logo
01 / AI CLIENTS AND AGENTS

What Argent does with Claude Code

Every call Claude Code makes to a model goes through Argent first. Argent knows which agent sent it and which person owns that agent, checks it against your policy, and writes a signed record of what it decided.

  • A named agent with a human owner. Each install of Claude Code gets its own Argent agent identity tied to an accountable person, so no call is anonymous. Identity tethering
  • Only the models you allow. The agent can reach only the providers and models your policy names, and a request can be swapped to an approved model. Model control
  • Sensitive data handled before it leaves. Prompts and replies are scanned for personal and secret data, then redacted or blocked as your policy says. Sensitive data redaction
  • Spend priced and capped. Every request is priced and metered against budgets for the agent, its owner and the teams above them. AI spend caps
  • A signed record you can replay. Each decision becomes a hash-chained, signed audit record, and requests are grouped into sessions you can replay. Session replay
  • Tool actions checked before they run. Shell commands, file reads and writes, web fetches, MCP calls and prompt submits ask Argent first; if Argent cannot answer, the action is refused. Agent permissions
02 / AI CLIENTS AND AGENTS

How Claude Code connects

Claude Code reaches Argent in three ways.

  • TLS-inspecting forward proxy. Claude Code keeps its normal provider settings and sends its HTTPS traffic through the Argent forward proxy and trusts your Argent certificate; the exact settings are below. The proxy signs in with the agent's own key, so every call is attributed, then runs the same checks as the gateway. It opens only known AI hosts, and its certificate is limited to those hosts.
  • Tool hooks. Before Claude Code runs a tool or submits a prompt, its hook asks Argent for a decision and a data-loss scan. If Argent cannot answer, the hook refuses.
  • Gateway base URL. Claude Code points its model URL at your Argent route and uses an Argent agent key in place of the provider key. Argent removes that key and adds the real provider credential from your vault on the way out, so the tool never holds it.

One command does the setup. argent connect claude-code registers a per-tool agent under the signed-in person, keeps its key only in the OS keychain, and writes the tool's own config with a backup. --revert puts the old config back. Keyless agents explains why keys stay out of config files.

A small loopback relay holds the agent key and adds it to each proxy connection, so the key is never written into the tool's settings.

Proven with the real tool. Argent's team has driven Claude Code itself through a live Argent gateway and checked the signed records it produced.

The same steps are in the Argent console under Integrations > AI Clients, and each agent's Connect tab gives the copy-paste values for your tenant.

03 / AI CLIENTS AND AGENTS

What you configure

What your team sets in Claude Code. These are the settings the Argent console shows for this tool. Values in angle brackets come from your tenant.

  1. Sign in the CLI to Argent. argent login (device login, token stored in the OS keychain), then argent connect claude-code. Default mode is forward-proxy: starts a loopback relay (systemd --user unit / OS equivalent), registers a per-tool agent under the signed-in person, writes ~/.claude/settings.json.
  2. What connect writes into settings.json. apiKeyHelper -> argent key claude-code; env.HTTPS_PROXY -> http://argent-relay:<secret>@127.0.0.1:<port>; env.NO_PROXY -> gateway host; env.NODE_EXTRA_CA_CERTS -> pinned Argent CA bundle; env.ENABLE_CLAUDEAI_MCP_SERVERS=false; env.CLAUDE_CODE_ENABLE_TELEMETRY=1 + OTEL_EXPORTER_OTLP_ENDPOINT=<gateway>/_argent/otlp (OTLP http/protobuf); otelHeadersHelper; hooks PreToolUse, UserPromptSubmit, UserPromptExpansion -> argent hook claude-code.
  3. Manual alternative (no argent CLI). Set env HTTPS_PROXY=<Argent forward proxy URL via a loopback http relay since Claude Code needs an http:// proxy URL> and NODE_EXTRA_CA_CERTS=<per-tenant intercept CA PEM>. Claude Code authenticates to Anthropic with its own key/OAuth; Argent sees the request at api.anthropic.com.
  4. Undo. argent disconnect claude-code restores settings.json byte-for-byte and revokes the agent.
04 / AI CLIENTS AND AGENTS

Roll it out to every machine

Make it the only way Claude Code runs on company machines. These controls are documented by the platform each one belongs to (linked); pair them with the Argent settings above.

  • Managed settings file (endpoint-managed) - Argent can emit it (macOS). Highest-precedence Claude Code config the user cannot override with ~/.claude/settings.json, project settings or --settings. Put env.HTTPS_PROXY, env.NODE_EXTRA_CA_CERTS (and optionally ANTHROPIC_BASE_URL), apiKeyHelper, hooks and allowManagedHooksOnly in it. Vendor docs
  • Managed settings file (Linux). Same as macOS (also WSL). Vendor docs
  • macOS MDM managed preferences (Jamf / Intune / Kandji configuration profile). Same settings JSON delivered as a profile; top-level settings are plist keys, nested objects dictionaries, lists arrays. Vendor docs
  • Windows registry policy (GPO / Intune custom OMA-URI / ADMX-less registry). Settings JSON as one registry value. Vendor docs
  • Server-managed settings (Claude.ai admin console). Org-wide JSON delivered at startup, polled hourly, for Claude.ai Teams/Enterprise sign-ins. Vendor docs

Argent ships the push kit. For managed fleets the console produces per-MDM scripts (Jamf, Kandji, Intune-style, shell and PowerShell) with your tenant certificate and proxy address built in, plus a fleet setup script and certificate bundle. Your company CA can sign the Argent certificate, so managed machines already trust it.

argent connect claude-code --emit-managed writes an admin-deployable managed-settings.json that pins the Argent key helper above user and project settings.

Want to see which AI tools people already run before you push anything? Start with Shadow AI discovery in watch mode. Watch before you enforce

05 / AI CLIENTS AND AGENTS

Check that it works

Send one request and look for the signed row. argent doctor claude-code (checks CA pin, session, keychain, relay running, settings.json mode 0600), then run claude -p 'say OK' and look for a signed llm ALLOW row with metering for agent argent-claude-code-<host> in the Argent console live stream; a canary SSN prompt should show a signed deny row.

Every request also appears on the live stream with the agent's name and owner, and in the session view. Session replay