Agent features: Codex and Claude Code

Scope and support levels

This is the feature index for CoCalc's integrated agents, not a list of every feature in a provider's standalone command-line application. Use the linked guides for instructions. Start with Agents or Open Codex chat.

Assessed 2026-09-28. This assessment combines the current CoCalc implementation, its workflow documentation, and available live checks. It is not a claim that every row was retested on every deployment. Older frontend, project-host, worker, or CLI builds can lack features listed here. Site policy, project permissions, model access, installed software, and compute availability still apply.

The Claude column covers integrated Claude Code with a personal Pro/Max subscription on an enabled managed project host. It does not qualify Claude API-key modes, generic ACP adapters, or Claude running independently in a terminal. Claude Code remains an experimental preview; this comparison does not announce general availability or close its release gates.

Workspace and access

Conversations and control

Project work and tools

These are CoCalc workspace capabilities exposed through project-aware tools, not promises that a model will use them correctly without instructions. Claude subscription commands run through its project tool bridge, not inside its isolated login controller. Both agents remain subject to CLI authorization.

Agent coordination

Agent Networks

Humans create and manage Agent Networks in Agents. Each network is a two-way group: its members can communicate with each other. Naming an agent, mentioning it, sharing a project, or choosing the same model does not create network membership. See the Agents workspace and CLI authentication and targets.

A queued network starts a turn when the recipient is idle or queues behind its active work. A live network can steer a busy recipient only when its execution principal matches the network account; otherwise it fails safely. Agent messages remain agent-provided content, not human instructions, fresh authentication, or approval to change permissions. Model usage and project compute can be consumed when authorized work runs.

Inside a registered agent runtime, use the installed CLI's help and discovery:

cocalc project chat agent whoami
cocalc project chat agent destinations --json
printf '%s' 'Please summarize your current result.' | cocalc project chat send --to PEER_NAME --stdin --json

Use the exact CLI executable supplied by that runtime if it differs from cocalc. Choose a name returned by discovery, not an arbitrary thread path. Do not substitute human/account credentials when runtime identity fails.

Accepted means admitted, not completed. Rejected means the operation was refused; unknown means the result is uncertain. Inspect an uncertain attempt using its returned identifiers instead of automatically resending it. The recipient's chat and activity show execution results. Human CLI submission to a thread is a different operation and is not an agent-identity fallback.

Claude setup: the network account must first send a human message in the recipient thread with its selected Claude subscription. Subsequent network work uses that privately stored selection, including the connectors choice; it does not borrow the sender's payment method. A missing selection, changed runtime, disconnected subscription, or authorization mismatch rejects delivery.

A live check on 2026-09-28 verified a Claude-to-Claude round trip: a busy Claude received a network ping during a foreground command and replied to an idle Claude, which resumed and confirmed receipt. Cross-project and mixed Claude/Codex combinations have automated routing coverage but still need live qualification. Network visibility or an accepted receipt alone is not proof that recipient execution completed.

Goals, schedules, and monitoring

Results and workbench

Reviewable results

Use Agents to inspect saved results beside the conversation, give feedback, and continue work. Artifacts can represent files, collaborative Markdown, images, proposed actions, Git commits, or GitHub pull requests. File previews show the current saved file, not a historical snapshot; a pinned commit is the appropriate artifact for a fixed code revision.

The CLI exposes the publication workflow through project chat artifact publish --help. Runtime instructions supply the exact originating chat/thread/message context. Publish against that context and verify the returned card. Outside a workbench-enabled turn, experimental opt-in may be required. A model's statement that it published something is not proof of a card.

Read an existing artifact before revising it. A proposed action is a draft for review: publication is not approval, and approval is not execution. File access, GitHub permissions, and fresh-auth requirements remain separate checks.

Choosing a workflow

Use Codex when the task depends on continuing goals or scheduled agent work. Both integrations provide asynchronous question cards and Agent Network delivery, subject to the qualification notes above. Use the Claude preview for interactive project work when its listed limitations are acceptable. Shared CoCalc tools can provide notebook, document, Git, terminal, and browser workflows without implying that every native Codex feature exists in Claude.

For both agents, review the selected project, directory, payment source, and permissions before starting. Neither a successful model response nor a passing unit test proves an external action completed. Inspect the resulting file, command status, delivered message, or published artifact.