CoWork's Agent Card

CoWork can describe itself with an A2A 1.0 agent card: what this CoWork can do and how its tools behave, so another agent, a teammate or an admin can see that before handing it work. CoWork already reads other agents' cards for PACT; this is the card it gives out.

What the card says

• Name, version and description. The description says CoWork runs on its user's machine and does not take tasks over A2A. The card lists no A2A interface (supportedInterfaces is empty), so an A2A client will not try to send it work.
• Skills: the profile's enabled task skills (id, name, description, category as the tag), up to 200.
• Tool behaviour, in an optional extension (urn:cowork-os:agent-card:tool-hints:v1, marked required: false, so strict card readers accept it):
  • builtInTools: every built-in tool with readOnly and destructive from CoWork's tool risk levels, reversible (true for read-only tools, false for destructive ones, null when unknown), and a costClass (local, or network for tools that call the network).
  • connectors: each enabled MCP connector and its tools, with the hints the connector itself publishes (readOnlyHint, destructiveHint). A hint the connector does not state is null, not a guess. Connector tool calls have the cost class external.
  • model: the default model's price class: local (no per-token price), metered (with its USD price per million input and output tokens) or unknown.

Seeing and publishing it

• cowork agent-card prints the card for this profile. It reads the local skills, the MCP settings (tools cached at each connector's last connection) and the provider settings.
• Publishing is off by default, because the card names the installed skills and connectors. cowork agent-card publish on sets the Control Plane setting publishAgentCard; the Control Plane (desktop app or coworkd) then answers GET /.well-known/agent-card.json with the current card (live connector tools when connected), Cache-Control: max-age=300 and Access-Control-Allow-Origin: *. While it is off, the path answers 404. Only GET and HEAD are allowed.
• The Control Plane binds to 127.0.0.1 by default, so the card is only reachable from this machine unless you expose the Control Plane yourself (for example with Tailscale). CoWork never publishes it anywhere on its own.

Estimating a task before running it

cowork tasks plan --estimate "task prompt" [--model <id>] gives a prepared-change-style estimate (kind: "summary", summary, estimatedCost in USD, estimatedDurationMs, and a payload with the sample sizes and the 90th percentiles). It is based only on this profile's last 30 tasks on the model: the cost is their median total, and the duration is the median wall-clock time of the completed ones, approval waits included. The prompt itself is not analysed. A local model has no per-token cost; with fewer than 3 past tasks, the estimate says there is not enough history.