Agent entrypoint

Give any agent the same work.

One command connects Claude Code, Codex or any MCP host to the tasks you already track. The host fixes the executable, space, actor and capability set before the model ever sees a tool.

connect an agent

tasq agent install claude --space my/project --actor claude:main

It prints the exact registration it would write, and only applies it with --apply. The manual per-host forms are below if you prefer to write them yourself.

Read SKILL.md Machine contract

Try the real package first.

Both runners execute the exact scoped package version without creating a persistent Tasq installation. The unrelated unscoped npm package is not Tasq.

try without installing

npx @tasq-run/cli@0.4.2 demo
bunx @tasq-run/cli@0.4.2 demo

Copy the host recipe. Replace every explicit value.

These recipes start the same local stdio MCP server. Use an absolute executable path. A space name on two isolated stores does not create coordination.

Codex

codex mcp add tasq -- \
  /absolute/path/to/tasq mcp \
  --tenant robotics/team-a \
  --actor codex:gwendall \
  --capabilities read,propose,coordinate

Claude Code

claude mcp add tasq --scope user -- \
  /absolute/path/to/tasq mcp \
  --tenant robotics/team-a \
  --actor claude-code:gwendall \
  --capabilities read,propose,coordinate

Generic MCP JSON

{
  "mcpServers": {
    "tasq": {
      "command": "/absolute/path/to/tasq",
      "args": [
        "mcp",
        "--tenant", "robotics/team-a",
        "--actor", "agent:gwendall",
        "--capabilities", "read,propose,coordinate"
      ]
    }
  }
}

A project file is a pointer, never permission.

The versioned rendezvous descriptor can name the Local store reference, space and agent instructions. It contains no token, credential or effect authority, and Tasq never activates it merely because it exists in the current directory.

Example descriptor Read the agent guide