Skip to content
Open Source Claude Cowork

Claude Cowork security: an open-source audit trail you own

The open-source answer to "can you see what the agent did?" starts with a fact Anthropic publishes: its audit logs do not record Claude Cowork activity, so the record is a transcript feed rather than a log you keep. Kortix is the open-source CompanyOS, my pick, and here the record is the place the agent ran; I work at Kortix, so weigh that accordingly. Checked 24 October 2026.

The audit gap, in the vendor's own words

Anthropic documents the gap itself. In the security review step of its Claude Cowork enterprise administrator guide, one line reads: "Audit logs record the events listed in Access audit logs, and Claude Cowork activity isn't among them; use the Compliance API for Cowork session transcripts." (Claude Academy, checked 24 October 2026.)

Harmonic Security's practitioner's guide, last updated 14 September 2026, pairs the two facts and calls the first one a correction: "The Compliance API now covers Cowork. The audit log still does not. This reverses our earlier guidance." The same page adds that "The Audit Log CSV export contains no Cowork event types", and that the Activity Feed "is not retroactive". A team that enables the Compliance API after rollout has a record only from that day forward (Harmonic Security).

What three gateway vendors claim, and where it narrows

Three vendors that sell gateways put the exclusion at its widest. TrueFoundry writes that "Cowork activity is explicitly excluded from Anthropic's Audit Logs, Compliance API, and Data Exports on every plan tier, including Enterprise." Mintive's article states "Cowork activity is explicitly excluded from Anthropic's Audit Logs, Compliance API, and Data Exports." Repello AI writes that "Anthropic's native compliance tooling (Audit Logs, Compliance API, Data Exports) does not cover Cowork activity on any plan tier" (TrueFoundry, Mintive, Repello AI).

Read those three beside the vendor's own page and the finding narrows to something you can check. The audit log gap is real and Anthropic states it. The Compliance API half is behind the product: Anthropic and Harmonic both describe a transcript feed for Enterprise organizations, and the three write-ups describe the state before that feed shipped. What survives all four sources is the part that matters in a review. There is no audit log entry for an agent's action, and no export a team can hold after the contract ends.

The remedy they all propose, and the interest behind it

All three vendors propose the same compensating control: route OpenTelemetry to your own SIEM. TrueFoundry is direct about what that feed carries, saying that "by default, prompts, MCP server names, tool names, and skill names are excluded from those logs" (TrueFoundry). Event-level metadata is not the same record as the command that ran and the file it touched.

Their finding and their remedy come from the same company. TrueFoundry sells an AI gateway and an MCP gateway, Repello sells an MCP gateway, and Mintive's article recommends MCP gateway visibility. That does not make the observation wrong. The audit log gap is on Anthropic's own page. It does mean the fix was never going to be a proxy. The record it hands you still sits on somebody else's side of the wall.

The defaults, and the accounts the question now reaches

The defaults are the part a reviewer should check first. Harmonic records that on Team plans Cowork, cloud sessions, Claude in Chrome and the built-in browser all start enabled. It also records that "Allow Automatically approve mode" is on by default, which puts a classifier rather than a person in the reviewer's seat before an action runs. On Enterprise the two browsers were off at launch and then "ON by default from 10 September 2026 unless already disabled", with no notice to members. A team that assumed the safer default was in place is running the other one.

The merge widened who this question applies to. On 16 September 2026 Anthropic folded Cowork into Claude: agentic work now happens inside ordinary conversations, and the Cowork toggle disappears as the rollout reaches each account (VentureBeat). A security team cannot block "Cowork" as a separate product any more and call the review closed.

Why a gateway is not an audit trail

A gateway records the traffic it can see. It sees a call leave for a model endpoint or an MCP server, and it can hold that call or log its arguments. The file the session opened before the call never crosses it. Neither does the shell line that produced an upload, or the scheduled task that ran at three in the morning. TrueFoundry's guide says where its own coverage stops. Local conversation history and file operations are covered by "your endpoint security posture", meaning disk encryption and EDR, and they never reach the proxy (TrueFoundry).

A gateway is also a second product to govern. It has its own deployment, keys, admin surface and uptime, and each of those is another thing to review. When it is down the trail has a hole; when it is misconfigured the hole is silent. The team ends up auditing a proxy's access policy next to the agent's, and the original question has doubled instead of closing.

A gateway can record that an action happened. It cannot record who approved it, because the approval is a click inside the vendor's client, made by the member in the session. In the automatic mode Harmonic describes, there is no click at all and a classifier decides, so the event arrives at the SIEM with no name attached to the decision.

When the authority for an action is a setting inside the product, the evidence of that authority is inside the product too. Anything in front of it can witness the action. The place where the agent runs can show the permission that allowed it, and the person who let it through.

What an auditable agent looks like instead

An auditable system starts with where the permissions are written. In Kortix an agent is one agents.<name> block in kortix.yaml: its prompt or prompt file, its model, mode and permission, and what it may reach. The reference is blunt about the default: "Omitted means none." That covers skills, subagents, commands, extensions, connectors, secrets, apps and kortix_permissions; tools defaults to "*" with a kortix validate warning, permission to allow, and codemode to on (the kortix.yaml reference). A permission change is a diff a colleague can read.

Access fails closed by default. The guide says what that means in practice: "Access is denied by default. An agent gets no connector, secret, skill, App or project repository until you grant it." (the agents guide). Access set in a console is a checkbox somebody ticked. Access written in a repo is a line somebody wrote and somebody else merged.

A two-minute run of Kortix, agent permissions as files, shows the fail-closed rule at a terminal: kortix init northwind writes the project, a grant swapped for a permission that does not exist makes kortix validate fail at the exact path naming it, and the corrected file validates clean.

The session is a branch of the repository. The starter agent's body describes the workspace it runs in as "an isolated sandbox", "an ephemeral branch of this repo, your own /workspace", where "Only what you commit + push survives". Each session gets one isolated sandbox with its own isolated machine and branch. What it writes and what it runs show up as repository state that a diff can reach.

Work then lands the way code lands. Session work reaches the default branch through a change request, and merge is default-deny for agents. A person holding the project.gitops.merge permission reviews the diff and merges it. The name on that merge is the approval, and it sits in the history beside the change it approved. An auditor who asks who authorized something gets a person, a diff and a timestamp.

The record is the repository. Commits, change requests, reviews and merges live in it, which is why a rename, a plan change or a vendor decision cannot take them away. Kortix runs on Kortix Cloud, in your VPC or on your own on-prem network, and self-host is free. The repo you clone holds the same history as the one you audit. Where the data sits is covered in Claude Cowork data privacy. How sessions run is on the self-hosted page, and what the product is in one piece is on whether Claude Cowork is open source.

A ten-minute review you can run on any product that runs agents

A reviewer can put four questions to any product that runs agents, and each one has an answer that can be shown. The fourth carries most of the weight: a settings panel leaves no file to diff or to keep.

  • Where permissions are written. Ask for a permission change as a diff, and ask what an omitted grant resolves to. A screenshot of a settings panel is not a diff, and "the admin decides" is not an answer about a file.
  • What the log is. Ask for an export the team keeps after the contract ends, carrying the command, the files touched and the time. If the answer is telemetry, ask whether it carries prompt content by default; on Cowork it does not.
  • Who approves an irreversible action, and what the default is. Ask what state a new workspace starts in, and whether a person or a classifier reviews. Harmonic's note on Anthropic's own telemetry is the reason to ask: people approved roughly 93% of permission prompts, so an approval nobody reads is a formality.
  • What a session can reach. Ask whether session work can reach the default branch without a person, and what the session can touch on the network. The answer you want is default-deny, with the grant named and the file it lives in. The four-product table runs those questions, and adds two rows a reviewer asks for: self-hosting and model choice.

Four products, seven questions

Isolation, permission files and an audit trail: what a reviewer can read unaided.

  1. Open source

    YesYes, the whole platformKortix
    • Claude, with Cowork inside itNoNo
    • OpenWorkYesYes
    • ChatGPT WorkNoNo
  2. Session isolation

    One isolated sandbox per sessionKortix
    • Claude, with Cowork inside itCloud by default on every plan; local desktop is a Team/Enterprise opt-out
    • OpenWorkThe member's machine; cloud workers optional
    • ChatGPT WorkA local sandbox or a Work Cloud session
  3. Where permissions are written

    One agents.<name> block in kortix.yamlKortix
    • Claude, with Cowork inside itCustom roles and capabilities in the admin console
    • OpenWorkAgent permissions and MCP tool policy in the organization
    • ChatGPT WorkA control under the composer; admins set policy in Agent Security
  4. Who approves an irreversible action

    A person merges the change requestKortix
    • Claude, with Cowork inside itAuto on Pro, Max, Ultra and Team; Manual asks first on Enterprise
    • OpenWorkTool policy per connection, set by the organization
    • ChatGPT WorkThe member, through Ask for approval
  5. Audit trail

    Every action is recorded. Reading, exporting and streaming the audit log depend on the planKortix
    • Claude, with Cowork inside itCompliance API transcripts; no Cowork events in the audit log
    • OpenWorkEvery agent action logged and exportable to a SIEM
    • ChatGPT WorkCompliance API events; not every local action is recorded
  6. Self-host

    Kortix Cloud, your VPC, or on-prem. Free.Kortix
    • Claude, with Cowork inside itNoNo
    • OpenWorkSelf-host or a managed private instance
    • ChatGPT WorkNoNo
  7. Model choice

    Any model provider, your own keysKortix
    • Claude, with Cowork inside itClaude models on a plan, or a configured provider
    • OpenWork50+ providers, or local models
    • ChatGPT WorkThe models OpenAI ships

One agent's permission rules, tool by tool

kortix.yaml
permission:bash:"*": allow"git push *": ask"rm -rf *": denyedit:"*": allow"secrets/*": deny

One agent's permission block: allow runs the call, ask holds it for a person, deny never runs it, and each tool can carry its own answer per glob.

Questions a reviewer asks first

Short answers, each sourced to the vendor page it came from.

  • Is Claude Cowork activity captured in Anthropic's audit logs?

    No. Anthropic's administrator guide says its audit logs record the events listed in Access audit logs, "and Claude Cowork activity isn't among them"; it points admins to the Compliance API for Cowork session transcripts. Harmonic reports that the Audit Log CSV export contains no Cowork event types. Enterprise organizations can retrieve transcripts through the Compliance API, from the day it is enabled.

  • Is Claude Cowork safe to use in a regulated industry?

    Only with Enterprise controls in place, and those still leave the audit gap open. There is no audit-log entry for a Cowork action yet, and automatic approval puts a classifier in the reviewer's seat. A regulated team needs Compliance API transcripts enabled before rollout, custom roles that scope Cowork per group, and read-only tools wherever a write is not required.

  • How do I audit what an AI agent did on my machine?

    From the agent's own log plus your endpoint record, or from the repository if the agent works in one. Ask the product for an export carrying the command, the files touched, the time and the approver. Where it keeps no such log, telemetry into a SIEM gives events, not content.

  • What is the difference between Claude Cowork and Claude Code from a security perspective?

    Cowork was the same agent architecture pointed at files, browsers and connectors instead of a repository, and since 16 September 2026 those abilities run inside ordinary Claude conversations. Claude Code stays a separate product for developers working against a codebase from a terminal or an IDE.

  • Can you self-host Claude Cowork?

    No. There is no server to install. Cowork stays Anthropic's app: a hosted service on a Claude plan, or a managed device in the third-party deployment mode, where inference and billing go to a provider the organization configures (the 3P overview). The Bedrock, Gemini Enterprise Agent Platform and Foundry routes change who bills the inference, not who operates the software. Kortix is open source, self-hosts on your own network, and self-host is free.

opensourceclaudecowork.com

© 2026

Put the record where you can read it

Agents, their permissions and the merge gate live in a repo you can clone today.