Skip to main content

Swif AI Agent Activity Monitoring Policy

The Swif AI Agent Activity Monitoring Policy records the sessions and tool activity of AI coding agents on managed macOS, Windows, and Linux devices, so admins can review what each agent did.

This policy was previously called the Swif ADR Sensor Policy.

It uses the open-source Uber Agentic Detection and Response (ADR) Sensor. The sensor reads the local logs that supported AI agents write, converts the activity into a common format, and uploads it to Swif. Use it to:

  • see which AI agents, models, and tools are used on your devices

  • review prompts, responses, and tool calls for a session

  • investigate suspicious commands, prompt injection, unauthorized file access, or possible data exfiltration

  • optionally have Swif analyze sessions for risky behavior

  • control which AI agents and tools may be used, and report or block shadow AI: agents your organization hasn't approved

  • find which account each AI agent is signed in to, and catch personal accounts used on company devices

  • control what AI coding agents can do before they act, such as which domains they can reach, with Action Authorization

Important: This policy collects sensitive data. Session records can include prompts, responses, project paths, tool arguments and results, usernames, and hostnames. Review Privacy and security considerations below and inform affected users before you turn it on.

Activity monitoring itself doesn't block anything. With Agent Access set to block, the policy also blocks agents and tools you don't allow, in Claude Code, Codex, and Cursor. To block destructive commands, use the Swif AI Agent Command Protection Policy.


Requirements

Requirement

Details

Operating systems

macOS, Windows, and Linux. Not available yet on Intel Macs or Windows ARM64.

Device ownership

Company-owned and BYOD

Agent

Swif Agentic Security installed and running. Version 1.28.0 or later is recommended for collection intervals under 20 minutes. Deploy it with the Apple, Windows, or Linux Agentic Security Policy.

Local data

At least one supported AI agent must have local session data on the device

Signed-in accounts

Swif Agentic Security agent 1.44.0 or later, with Collect Signed-in Accounts on

Agent Access for Hermes, Grok Bot, and OpenClaw

The Swif Agent Inventory Policy, assigned to the same devices and enabled, with the agent in its Agents list. Agent Access sees these agents only through the agent inventory.

Connectivity

The device must be online to receive the policy and upload data


How it works

  1. A supported AI agent writes session, conversation, or tool-use data to its local storage.

  2. On the schedule you set, the Swif agent runs the sensor, which reads that data.

  3. The sensor converts the activity into session records, covering the number of days you choose.

  4. If Upload Enabled is on, the records are uploaded to Swif.

  5. If Risk Detection Enabled is on, Swif analyzes the uploaded sessions for risky behavior.

  6. You review sessions in Agentic Security > Devices > [device] > Activity > AI Agent Sessions.


Settings reference

Setting

What it does

Default

Range

Enabled

Turns activity collection on or off. While it's off, nothing is collected or uploaded.

Off

—

Monitored Agents

The AI agents to monitor. Leave empty to monitor all supported agents.

Empty (all)

See below

History Window (Days)

How many days of session history to include in each upload.

14

1–30

Collection Interval (Minutes)

How often the device collects AI agent activity.

15

5–1,440

Upload Enabled

Uploads collected activity to Swif. When off, records stay on the device.

On

—

Risk Detection Enabled

Analyzes uploaded sessions for risky behavior. Requires Upload Enabled. Sends session data to an approved external LLM provider.

Off

—

Collect Signed-in Accounts

Reports the email, organization, and plan each AI agent is signed in to, so you can find personal accounts on company devices. Tokens are never collected. See Signed-in accounts below.

Off

—

Agent Access

Which AI agents, and which of their built-in tools, may be used. Disallowed use is reported as shadow AI, and blocked where the agent supports it. See Agent Access: control shadow AI below.

Off

—

Action Authorization

Allows, blocks, or asks approval for each action an AI coding agent tries, such as reaching a domain or running a command. See Action Authorization below.

Off

—

If you turn on Enabled and keep the other defaults, the device collects from all supported agents every 15 minutes, includes up to 14 days of history, and uploads it to Swif. Risk detection stays off.

Monitored agents

Value

AI agent

macOS

Windows

Linux

Local data read

claude

Claude Code

✅

✅

✅

JSONL session files

codex

OpenAI Codex CLI

✅

✅

✅

JSONL session files

claude_desktop

Claude Desktop local agent mode, including Claude Cowork

✅

✅

—

Local audit JSONL files

cline

Cline

✅

✅

✅

Local JSON task files

cursor

Cursor

✅

✅

✅

Local SQLite app state

warp

Warp Terminal

✅

✅

—

Local SQLite app data

hermes

Hermes Agent

✅

Not yet

✅

Each user's Hermes session store, ~/.hermes/state.db. Swif reads a copy, so the live database isn't touched.

Agents that aren't available on a device's operating system are skipped. Hermes needs a Swif agent version with the new activity monitoring runtime, and isn't yet supported on Windows.

Note: Warp sessions come only from Warp's own AI agent. Warp Terminal shows up only when you use Warp's built-in agent. If you run another AI agent inside Warp, such as typing claude to start Claude Code, the session is reported under that agent (for example, Claude Code), not Warp. Warp's agent also needs a signed-in Warp account. If the user skipped the Warp login, Warp's AI doesn't run, and there are no Warp sessions to collect.

Note: Claude Code is monitored in the terminal or an IDE. Claude Code's extensions and plugins for IDEs, such as VS Code, run the same Claude Code engine as the command-line tool. Their sessions are recorded as Claude Code sessions, along with the MCP servers they're connected to. It doesn't matter how Claude Code signs in: with a Claude account, an API key, Amazon Bedrock, or Google Vertex AI.

Direct API use isn't recorded. This policy reads the session files that supported agents, such as Claude Code, save on the device. If a developer's own script or tool calls the Anthropic API directly, without Claude Code, there are no session files, so nothing is recorded.

Leaving Monitored Agents empty collects from every supported agent. It also includes opencode, which isn't available as a separate option. To limit collection, list only the agents your organization has approved.

Choosing the history window and interval

  • History Window (Days): a shorter window means less data and less privacy exposure. A longer window gives more context for investigations. It doesn't delete data already uploaded to Swif. Apply your own retention rules to that data.

  • Collection Interval (Minutes): a shorter interval shows sessions in Swif sooner, but uses more processing and network. The 15-minute default suits most organizations. Use 5 minutes only for an active investigation.

Collection timing depends on the agent version:

Agent version

How the interval works

1.28.0 and later (macOS, Windows, Linux)

Collection runs on its own schedule, checked every minute. Intervals from 5 to 15 minutes take effect as set.

1.27.0 and earlier

Collection runs with the agent's 20-minute policy check. Any interval under 20 minutes effectively runs every 20 minutes.

On any version, a policy change can take up to 20 minutes to reach the device.


Data collected

What's collected depends on the AI agent, its version, and what it records locally. It can include:

Category

Examples

Device and user

Hostname and local username

Session

Session ID, timestamp, and agent

Model

Model name recorded by the agent

Project

Local project or workspace path

Conversation

User prompts and assistant responses

Tool activity

Tool name, type, arguments, results, and status

Agent configuration

Available tools, MCP servers, plugins, skills, or permission mode, where the agent records them

The sensor only reads what the agent already logged. It can't collect activity the agent didn't record, or sessions the agent or user has deleted.


Risk detection

When Risk Detection Enabled is on, Swif analyzes each uploaded session for:

  • Credential theft and access: attempts to read local credentials, environment secrets, or unauthorized folders

  • Prompt injection: malicious instructions in code, issues, or web content that steer the agent

  • Data exfiltration: unapproved transfer of code or sensitive data through tool calls

  • Destructive operations: high-risk shell commands, system changes, or attempts to evade security controls

Each analyzed session gets a finding:

Field

Values

Verdict

Malicious, Benign, or Unknown

Analysis State

Complete, Partial, or Failed

Summary and reasoning

A short explanation of why the session was flagged

Risk detection sends full sessions to an external AI provider. To score a session, Swif sends the full session, including prompts, responses, and tool inputs and outputs, to an approved external LLM provider. Analysis runs on Swif's servers, never on the device. Findings contain only the verdict, summary, and reasons, not transcripts or tool output, and paths and secrets are redacted from the summary.


Signed-in accounts

Employees sometimes use AI agents on work devices while signed in to personal accounts, such as a personal Gmail address on Claude, or a personal Cursor plan. Company code and data can then end up in accounts your organization doesn't control.

With Collect Signed-in Accounts on, Swif finds the account each AI agent is signed in to, for every user on the device. It reads the agents' local sign-in files, not their sessions, so agents that are installed but haven't been used yet are found too.

Only the email, organization, and plan are collected. Tokens, passwords, and other credentials are never collected. As a second safeguard, Swif rejects any value that looks like a token. The full email address is stored and shown to admins.

Supported agents

Agent

What's reported

Claude Code, Claude Desktop, Codex, Cursor, Antigravity, Gemini CLI

Email, organization, and plan

GitHub Copilot

GitHub username, instead of an email address

Cline, Warp, Hermes, opencode

Not supported. These agents don't store a readable account.

OpenClaw, Grok Bot

Not supported yet

Accounts are collected for every supported agent, whatever you choose in Monitored Agents. Monitored Agents only controls activity monitoring.

How accounts are classified

Classification

When it applies

Corporate

The email uses your organization's domain, or the account is on a business or team plan.

Personal

The email uses a consumer domain, such as gmail.com, or the account is on a personal plan, such as Claude Max or Cursor Pro.

Unknown

There's no email (for example, Copilot uses a GitHub username), the agent isn't signed in, the agent isn't supported, or the email's domain isn't recognized.

How Swif knows your organization's domain. Swif works it out from the email addresses of the people enrolled in your workspace. For example, if your users enroll with @acme.com addresses, acme.com is treated as your organization's domain. Public email domains, such as gmail.com and outlook.com, are never counted.

Changes over time

Swif records when an account is signed in, signed out, or switched, and when its classification changes. For example, it records Cursor switching from jdoe@acme.com (Corporate) to jdoe@gmail.com (Personal).

Each AI agent session also records the account it ran under, so you can see which identity was used for a session.


Agent Access: control shadow AI

Shadow AI is AI agents or tools that employees use without your organization's approval. They can read code, run commands, and send data to outside services, with no oversight. Agent Access lets you decide which AI agents, and which of their tools, may be used on your devices.

Mode

Mode

What happens

off (default)

Nothing is checked.

report

Disallowed agent and tool use is detected and reported, but not blocked.

block

Disallowed use is also blocked, on agents that support blocking. Everything else is reported.

Start with report. Review what your teams use, adjust your lists, then switch to block.

Agent list

Setting

What it does

Default

Agent List Mode

allow: only the listed agents may be used, and any other agent is shadow AI. deny: the listed agents may not be used.

deny

Agents

The agents the list mode applies to

Empty

Up to 32 agents. Use these values:

Value

Agent

claude

Claude Code

codex

Codex

cursor

Cursor

claude_desktop

Claude Desktop

cline

Cline

warp

Warp

hermes

Hermes

grokbot

Grok Bot

openclaw

OpenClaw

Allowlist or denylist?

  • Allowlist (allow): best when you've approved a short list of agents. Anything new is reported, or blocked, automatically.

  • Denylist (deny): best when most agents are fine, and you only want to stop a few.

Tool rules

Allow or deny specific built-in tools, for one agent, or for every agent. Each rule has:

Field

What to enter

Agent

An agent value from the table above, or * for every agent without its own rule

List Mode

allow: only the listed tools may be used. deny: the listed tools may not be used. Required, with no default.

Tools

Up to 16 tools, from the table below

Value

Tool

shell

Shell commands

file_read

Reading files

file_write

Writing files

file_search

Searching files

web_fetch

Fetching web pages

web_search

Web search

subagent

Starting subagents

notebook_edit

Editing notebooks

You can add up to 32 tool rules: one per agent, plus one * rule for all other agents. A policy can't have two rules for the same agent, or two * rules.

Swif checks the policy when you save it. If something is invalid, such as an unsupported tool name, the save fails, and the error names the field.

What can be blocked, and what's only reported

Agent

With block

Claude Code

Blocked on the device

Codex

Blocked on the device. The block is locked on in Codex's admin settings, so users can't turn it off.

Cursor

Blocked on the device, except web fetch and web search, which run on Cursor's servers and are only reported

Claude Desktop, Cline, Warp

Reported only. Detected from their monitored sessions.

Hermes, Grok Bot, OpenClaw

Reported only. Detected from the device's agent inventory, with the Swif Agent Inventory Policy.

Codex web search isn't visible to Swif, so it can't be reported or blocked.

What users see when something is blocked

Blocked

What happens

A denied agent

Every action the agent tries is blocked, with the message "Blocked by your organization's AI agent policy."

A denied tool

Only that tool is blocked. The agent can keep using its other tools.

What's recorded

Each shadow AI event records the device, its operating system, and the user. It also records the agent and tool, whether the use was blocked or reported, the reason (agent not allowed, or tool not allowed), how it was detected (hook, session, or inventory), the session ID, and the policy version that applied.

You can filter events by device, user, agent, tool, decision, reason, detection source, and date range, and group them into reports by any of these, or by day. Events are kept for 90 days.

Enforcement status on each device

Device state

Meaning

Disabled

Agent Access is off.

Reporting

Disallowed use is reported.

Enforcing

Disallowed use is blocked where supported.

Hooks incomplete

Blocking isn't fully set up for every agent on the device.

Policy invalid

The policy has an error.

Each agent on a device also has its own state: enforcing, reporting, report only, hook missing, or not installed.


Action Authorization

Action Authorization decides, before an AI coding agent acts, whether it may: run a command, change a file, fetch a URL, or call an MCP tool. Each action is allowed, denied, paused for an admin's approval, or allowed with secrets redacted. A denied action never runs.

Decisions can depend on the agent's trust state, the device's posture, the user, the tool, the action's risk, and what it touches, such as a domain or a file path.

Use an allowlist, not a blocklist. AI agents reach many services in normal use. Blocking a whole provider can break approved tools and developer workflows. Instead, allow the domains you trust, and deny, or require approval for, everything else.

Settings

Setting

What it does

Default

Mode

off: nothing is decided. report: every action runs, and the decision it would have had is recorded. enforce: decisions are applied.

off

Fail Mode

What happens when the device can't decide, for example when its rules or context can't be read, or are out of date. open lets the action run. closed denies it.

open

Context Max Age (Minutes)

How old the device's copy of agent trust, device posture, and approvals can be. After that, the device can't decide, and Fail Mode applies.

1440 (5–10,080)

Approval Duration (Minutes)

How long an admin's approval or rejection applies to the same action

60 (5–10,080)

Log Allowed Actions

In enforce mode, also records allowed actions. Denied, approval, and redacted actions are always recorded.

Off

Default Outcome

The decision when no rule matches

ALLOW

Rules

Up to 64 rules. The first rule that matches decides.

—

Mode names differ between Action Authorization and Agent Access. Action Authorization uses enforce to apply its decisions. Agent Access, in the same policy, uses block.

Outcomes

Outcome

What happens

ALLOW

The action runs

DENY

The action is denied, and never runs

REQUIRE_APPROVAL

The action waits for an admin's approval in Swif

ALLOW_WITH_REDACTION

The action runs, with matching secrets replaced by [REDACTED]. See Redaction.

Rules

Each rule has:

Field

What to enter

ID

Required. Unique in the policy: 1 to 64 characters, using a–z, 0–9, _, and -

Name

Shown in the decision log. Up to 128 characters.

Outcome

Required. What happens to a matching action.

Fallback

With ALLOW_WITH_REDACTION only: the outcome when the agent can't redact what the rule asks for: DENY (default), REQUIRE_APPROVAL, or ALLOW

Redaction

Required with ALLOW_WITH_REDACTION. See below.

A rule matches when every field it sets matches. An empty field matches anything. Each field takes up to 32 values.

Field

Matches

Values

Agents

The AI agent

claude (Claude Code), codex, cursor

Tools

The kind of action. mcp is any MCP server tool.

shell, file_read, file_write, file_search, web_fetch, web_search, subagent, notebook_edit, mcp

MCP Servers

MCP server names, for mcp actions only

Server names

Risks

The action's risk, as Swif classifies it

SAFE, RISKY, DANGER

Trust States

The agent's attestation trust state

TRUSTED, DRIFTED, UNTRUSTED, UNKNOWN

Device Compliance

Whether the device is compliant

compliant, noncompliant, unknown

Disk Encryption

Whether the device's disk is encrypted

encrypted, unencrypted, unknown

OS Users

The user's username on the device

Usernames

Paths

File paths. * stays in one folder, ** crosses folders, and ~ is the user's home folder.

Such as ~/.ssh/**

Domains

Domains

example.com matches only that domain. *.example.com matches any subdomain, but not example.com itself, so add both if you need both.

Commands

Shell commands, as RE2 regular expressions, up to 256 characters each

Such as curl .* | (ba)?sh

unknown means the device hasn't reported its compliance or encryption.

How Swif classifies risk

Risk

Actions

DANGER

Deleting files recursively, piping downloads to a shell, using disk tools, gaining privileges, reading credentials, writing to system folders, and changing permissions

RISKY

Network downloads, installing packages, pushing to Git, writing outside the project folder, fetching web pages, calling MCP tools, and starting subagents

SAFE

Everything else

Redaction

With ALLOW_WITH_REDACTION, matches are replaced with [REDACTED]:

Setting

What it does

Default

Redact Input

Redacts what the agent sends to the tool

—

Redact Output

Redacts what the tool returns to the agent

—

Built-in Secret Patterns

Redacts API keys, tokens, and passwords

On

Patterns

Extra RE2 expressions to redact. Up to 16 patterns, of up to 256 characters each.

Empty

If the agent can't redact what the rule asks for, Fallback applies.

What each agent can redact

Agent

Input (what it sends)

Output (what it gets back)

Claude Code

Every tool

shell, file_read, and mcp

Codex

shell and mcp

None

Cursor

shell, web_fetch, mcp, file_read, and file_search. Not the content of a file write.

mcp only

For example, a rule that asks Codex to redact output always uses Fallback.

Set up a domain allowlist

  1. Start in report mode. Set Mode to report, and assign the policy to a pilot group. Every action runs, and the decision log shows what would have been decided, so you can see which domains your agents reach.

  2. Allow approved domains. Add ALLOW rules, with Tools set to web_fetch, for the services you trust: your company's APIs, your approved AI providers, and your Git hosts. Add both example.com and *.example.com if you need the domain and its subdomains.

  3. Restrict everything else. Set Default Outcome to DENY, or add a last rule with DENY or REQUIRE_APPROVAL.

  4. Switch to enforce mode. Once the decisions look right, set Mode to enforce.

Order matters. The first matching rule decides, so put specific rules, such as allowed domains, before broad ones, such as "deny everything else".

Review decisions and approvals

  • Decisions: in report mode, every decision is recorded. In enforce mode, denied, approval, and redacted actions are always recorded, and allowed ones only with Log Allowed Actions on. Each shows the rule's Name.

  • Approvals: actions waiting for approval appear in Agentic Security > Shadow AI. An approval or rejection applies to the same action for Approval Duration (Minutes).

Limits

Limit

What it means

Supported agents

Claude Code, Cursor, and Codex

Vendor proxies

If an agent sends traffic through its vendor's own servers, instead of reaching a domain directly, Swif sees and controls only the vendor's address.

VPNs and encrypted DNS

A VPN or encrypted DNS (DNS over HTTPS) set up by the user can get around domain checks. Manage VPN and DNS settings centrally, for example with the Apple DNS Settings Policy.

Out-of-date context

If the device's context is older than Context Max Age (Minutes), Fail Mode decides. With closed, actions are denied until the device updates.

Action Authorization and other policies

Policy

Controls

Agent Access, in this policy

Which agents and built-in tools can be used at all

Action Authorization, in this policy

Each individual action, such as which domains and commands are allowed

Blocks destructive shell commands


Before you start

  1. Complete a privacy and security review, including whether risk detection's use of an external LLM provider is acceptable.

  2. Inform affected users that AI agent session data will be collected.

  3. Decide which AI agents to monitor. List only approved agents.

  4. Choose a pilot group of company-owned devices that covers each operating system in your fleet.

  5. Decide on BYOD. Don't reuse your company-owned configuration on BYOD devices without a defined business scope, employee notice, and an approved data-handling process.


Create the policy

  1. In Swif, go to Agentic Security > Policies.

  2. Select Swif AI Agent Activity Monitoring Policy.

  3. Enter a name, such as AI Activity Monitoring – Pilot.

  4. Turn on Enabled.

  5. Under Monitored Agents, select the approved AI agents.

  6. Set History Window (Days) and Collection Interval (Minutes).

  7. Turn Upload Enabled, Risk Detection Enabled, and Collect Signed-in Accounts on or off.

  8. Optionally, set Agent Access to control shadow AI. Start with report.

  9. Assign the policy to your pilot devices and save.

Expand to more devices after the pilot and your privacy, security, and legal reviews are complete.


Example configurations

Example 1: Pilot on company-owned devices

Setting

Value

Enabled

On

Monitored Agents

claude, codex, cursor

History Window (Days)

7

Collection Interval (Minutes)

15

Upload Enabled

On

Risk Detection Enabled

On, after privacy review

An explicit agent list and a 7-day window keep collection small, while giving enough recent activity to validate the policy.

Example 2: Check locally before uploading

Setting

Value

Enabled

On

Monitored Agents

codex

History Window (Days)

1

Upload Enabled

Off

Risk Detection Enabled

Off

Activity is collected but stays on the device. Use this during a technical or privacy evaluation. Sessions won't appear in Swif until you turn on upload.

Example 3: Active investigation

Setting

Value

Enabled

On

Monitored Agents

claude, claude_desktop, cursor

History Window (Days)

30

Collection Interval (Minutes)

5

Upload Enabled

On

Risk Detection Enabled

On

Collect the most history as often as possible. Return to the default interval and a shorter window when the investigation ends.

Example 4: Find shadow AI without blocking anything

Setting

Value

Enabled

On

Agent Access mode

report

Agent List Mode

allow

Agents

claude, cursor

Every other agent your teams use is reported as shadow AI, so you can see what's in use before you block anything.

Example 5: Approved agents only, with web access turned off

Setting

Value

Enabled

On

Agent Access mode

block

Agent List Mode

allow

Agents

claude, codex

Tool rule

Agent: *; List Mode: deny; Tools: web_fetch, web_search

Only Claude Code and Codex may be used. Other agents are blocked where possible, and reported otherwise. Neither approved agent can use web fetch or web search, except Codex web search, which Swif can't see.

Example 6: Find personal AI accounts on company devices

Setting

Value

Enabled

On

Collect Signed-in Accounts

On

Swif reports the account each AI agent is signed in to, and labels it Corporate, Personal, or Unknown. Use it to find employees using personal AI accounts for company work.

Example 7: Allow only approved domains

Setting

Value

Enabled

On

Action Authorization > Mode

enforce, after testing in report

Action Authorization > Default Outcome

DENY

Rule ID

Name

Match

Outcome

company

Company services

Domains: yourcompany.com, *.yourcompany.com

ALLOW

ai-providers

Approved AI providers

Domains: api.anthropic.com, api.openai.com

ALLOW

git

GitHub

Domains: github.com, *.github.com

ALLOW

danger

Dangerous actions

Risks: DANGER

REQUIRE_APPROVAL

Claude Code, Cursor, and Codex can reach only your company's services, your approved AI providers, and GitHub. Dangerous actions wait for approval. Everything else is denied.

Default Outcome applies to every action that matches no rule, not only web requests. With DENY, also add ALLOW rules for the commands and files your developers need, or set Default Outcome to ALLOW, and add a final DENY rule with Tools set to web_fetch.

Example 8: Redact secrets from shell commands

Rule ID

Match

Outcome

Redaction

redact-shell

Tools: shell

ALLOW_WITH_REDACTION, Fallback DENY

Redact Input and Redact Output on, Built-in Secret Patterns on

API keys, tokens, and passwords are replaced with [REDACTED] in shell commands and their output. Codex can't redact output, so with this rule, Codex shell commands are denied. To allow them, set Redact Output off, or Fallback to ALLOW.


Verify the policy

Use a test session that contains no credentials, customer data, or proprietary code.

  1. Confirm that the device is online and the policy shows as applied.

  2. Run a short session in one of the monitored AI agents.

  3. Wait for the policy to reach the device (up to 20 minutes), then for the collection interval to pass (15 minutes by default).

  4. Go to Agentic Security > Devices, select the device, and open Activity > AI Agent Sessions.

  5. Confirm that the session appears with the expected agent, model, project path, and time.

  6. Open the session and check the Chat History and Tool Usage tabs.

  7. Confirm that agents not in Monitored Agents don't appear in new sessions.


View AI agent sessions

  1. In Swif, go to Agentic Security > Devices.

  2. Select the device.

  3. Open the Activity tab, then AI Agent Sessions.

Search and filter

Filter

What it does

Search

Matches session ID, username, project path, model, or agent. Partial text works, such as sonnet or sample-repo.

Source

Filters by AI agent: claude, codex, cursor, cline, claude_desktop, or warp

Time range

Shows sessions that started between a start and end time

If the policy is active but no sessions have been collected yet, the list is empty.

Session details

Click a session to open its details.

Chat History (shown first) lists the messages between the user and the AI agent, labeled User and Assistant.

Tool Usage lists the tools, commands, arguments, times, and results from the session.

Each tab can show:

State

Meaning

Empty

The agent supports this data, but none was recorded in the session.

Unsupported

The agent doesn't record this kind of data.

Partial data (Tool Usage)

Some tool calls weren't fully captured in the agent's logs, as can happen with Codex CLI.


Privacy and security considerations

  • Tell affected users that AI agent session data may be collected.

  • Document the business purpose and legal basis for collection.

  • Check whether prompts, responses, tool results, or paths could contain secrets, personal data, customer data, or proprietary code.

  • List only approved agents in Monitored Agents.

  • Use the shortest History Window (Days) that meets your needs.

  • Limit who in your organization can view uploaded sessions in Swif.

  • Set retention and deletion rules for uploaded data.

  • Test BYOD devices separately from company-owned devices.

  • Review risk detection's use of an external LLM provider before turning it on.

  • With Collect Signed-in Accounts on, each AI agent's full account email is stored and shown to admins. Tell users before you turn it on, especially on BYOD devices.

The policy doesn't change the AI agents' own logging or privacy settings. Review each agent's storage, cloud processing, and retention separately.

The ADR Sensor is open source under the Apache License 2.0. Review its source code and release history as part of your software supply-chain process.


Change or turn off the policy

  • To change collection: edit the policy, save it, and let devices check in. A change can take up to 20 minutes to reach a device. The next collection after that uses the new settings.

  • To stop collection: turn off Enabled or unassign the policy. No new data is collected after devices check in.

Turning off the policy doesn't delete:

  • the AI agents' own session logs

  • records already created on the device

  • sessions already uploaded to Swif

  • data forwarded to other systems

Follow your retention and deletion procedures for each location.


Troubleshooting

The policy is applied, but no sessions appear

  • Check that Enabled and Upload Enabled are on.

  • Check that the device is online and the Swif agent is running.

  • Check that a monitored AI agent has recent local session data.

  • Check that the session falls within History Window (Days).

  • Wait up to 20 minutes for the policy to reach the device, then for Collection Interval (Minutes) to pass, and refresh.

Only some AI agents appear

  • Check that the missing agent is in Monitored Agents, or that the list is empty.

  • Check that the agent is supported on the device's operating system. Claude Desktop and Warp aren't supported on Linux.

  • Check that the agent has created local session data.

Claude Code sessions from an IDE don't appear

Sessions from Claude Code in an IDE, such as VS Code, are recorded as Claude Code sessions, not as a separate agent. Check that claude is in Monitored Agents, or that the list is empty. Then wait for the next Collection Interval (Minutes).

Calls to the Anthropic API don't appear

This is expected. This policy reads the session files that supported agents, such as Claude Code, save on the device. A script or tool that calls the Anthropic API directly, without Claude Code, creates no session files, so nothing is recorded.

Warp sessions don't appear

  • Check whether the session was run with Warp's own agent. If the user ran another agent in a Warp tab, such as Claude Code, the session appears under that agent instead, for example as a Claude row.

  • Check that the user is signed in to Warp (Settings > Account). If the login was skipped, Warp's AI can't run, and nothing is recorded locally.

To test Warp:

  1. Sign in to Warp.

  2. In a Warp tab, type /agent to open Warp's agent. Don't run claude or another agent's command.

  3. Ask something that runs a command, such as "list the files in this folder."

  4. Wait for the next Collection Interval (Minutes) to pass, then check Activity > AI Agent Sessions for a Warp row.

Older sessions are missing

Increase History Window (Days), up to 30. Also check that the AI agent still has the session locally. The sensor can't collect a session after its records are deleted.

Sessions are collected but don't appear in Swif

  • Check that Upload Enabled is on.

  • Check that the device is online and can reach Swif.

  • Check that you're viewing the correct device and time range.

If the device was offline, wait for it to reconnect and complete another collection.

Sessions arrive every 20 minutes, even with a shorter interval

The device runs Swif agent 1.27.0 or earlier, where intervals under 20 minutes effectively run every 20 minutes. Update the agent to 1.28.0 or later.

Collection uses more resources than expected

  • Increase Collection Interval (Minutes).

  • Reduce History Window (Days).

  • List only the agents you need in Monitored Agents.

A user sees "Blocked by your organization's AI agent policy"

The agent, or one of its tools, isn't allowed by Agent Access. If it should be allowed, add it to your allowlist, or remove it from your denylist.

A disallowed agent isn't blocked

  • Check that Agent Access is set to block, not report.

  • Only Claude Code, Codex, and Cursor can be blocked. Claude Desktop, Cline, Warp, Hermes, Grok Bot, and OpenClaw are reported only.

  • Check the device's enforcement status. Hooks incomplete or hook missing means blocking isn't fully set up on that device.

Hermes, Grok Bot, or OpenClaw use isn't reported

Agent Access sees these agents only through the agent inventory. Check that the Swif Agent Inventory Policy is assigned to the same device, enabled, and has the agent in its Agents list.

Hermes sessions don't appear on Windows

This is expected. Hermes session monitoring isn't yet supported on Windows.

Cursor web fetch or web search isn't blocked

This is expected. They run on Cursor's servers, so Swif can only report them.

Codex web search doesn't appear

This is expected. Swif can't see Codex web search, so it can't report or block it.

An agent shows "Not signed in"

The agent is installed, but no one is signed in to it. Accounts appear once a user signs in.

An agent shows "Not supported"

This is expected for Cline, Warp, Hermes, and opencode, which don't store a readable account. OpenClaw and Grok Bot aren't supported yet.

A GitHub Copilot account is labeled Unknown

This is expected. Copilot reports a GitHub username, not an email address, so Swif can't tell whether it's corporate or personal.

A company account is labeled Unknown

Swif doesn't recognize the email's domain as your organization's. Make sure your users are enrolled with their work email address. Swif works out your organization's domains from the email addresses of enrolled users.

Signed-in accounts don't appear

Check that Collect Signed-in Accounts is on. It's off by default.

An agent action was blocked unexpectedly

Check the decision for that action, and which rule's Name it shows. Rules are checked in order, and the first match decides. A rule matches only when every field it sets matches. If no rule matched, Default Outcome applied. Add an ALLOW rule above the one that blocked it, or test the change in report mode first.

A domain is denied, though I allowed *.example.com

*.example.com matches subdomains, such as api.example.com, but not example.com itself. Add example.com to the same rule.

A redaction rule denies actions from Codex or Cursor

The agent can't redact what the rule asks for, so Fallback applies. For example, Codex can't redact output, and Cursor redacts output only for mcp. See What each agent can redact. Turn off Redact Output for those agents, or set Fallback to ALLOW or REQUIRE_APPROVAL.

Every action is denied

The device's context may be out of date, with Fail Mode set to closed. Check that the device is online, and syncing with Swif. Or check Default Outcome: with DENY, every action that matches no rule is denied.

The same action keeps asking for approval

An approval applies to the same action only for Approval Duration (Minutes). After that, the action asks again.

A blocked domain is still reached

  • Check that Mode is enforce, not report.

  • The agent may send traffic through its vendor's servers, or the device may use its own VPN or encrypted DNS. See Limits under Action Authorization.

Sessions don't have a risk verdict

  • Check that Risk Detection Enabled and Upload Enabled are both on.

  • Check the session's Analysis State. Partial or Failed means the analysis didn't complete.


Related resources

Swif

Uber ADR

Did this answer your question?