Meta Harness for Power Users.

Connect Claude Code, Codex, Goose, or any ACP-compatible agent through a standard JSON-RPC transport. Swap providers without rewriting a single line.

Agents call task_spawn to create child tasks that route to new agent processes. Decompose complex goals into focused sub-tasks with automatic progress reporting.

Idle agents are swapped out to disk and restored on demand via ACP session/load. Run more concurrent long-lived tasks than there are live processes, just like OS virtual memory.

ACP reuses MCP types natively. Your existing MCP servers plug straight into each task sandbox, and a built-in MCP server exposes orchestration primitives out of the box.

Every ACP packet, in both directions, is logged as JSONL. Tool calls, permission decisions, and thinking traces are preserved for audit trails, debugging, and replay.

No Redis, no message broker. The filesystem is the single source of truth: task state, inbox messages, and MCP config all live as files, making deployment and crash recovery trivial.

A durable task phase and a volatile session phase are tracked independently. The agent process can crash, swap out, and restart, and the task picks up right where it left off.

Every message to an agent, including user prompts, sub-task results, and system events, flows through a write-ahead log with at-least-once delivery and deduplication, surviving even full session swaps.

MCP servers, agent runtimes, skills, and instruction files (AGENTS.md) are managed in one layered config that merges builtin, user, and workspace overrides. Profiles compose them all into a single deployable unit per task.

Agents run persistently on the server. Start a task from your laptop, close the lid, continue on your phone, or watch the same task from multiple tabs at once, all streaming live via SSE.

Tool calls block until the user explicitly approves, rejects, or auto-allows them. You get granular control over what each agent can do without modifying the agent itself.

Run local, SSH-remote, or reverse-connected workspaces from a single gateway. Each workspace gets its own process, admission queue, and idle-eviction policy.