Project status

Every capability on this site carries one of three labels:

• Built in the code on dev today.
• Partly built a working piece exists, but not the full idea.
• Vision not built. A direction, not a promise.

Phases

Phase	State
0. Spikes and kill criteria	Done (kill criteria not triggered)
1. Durable core	Done
2. Agents and trust	Done
3. Flagship workflow and alpha demo	Done. Checks 2 to 6 pass in npm run demo; check 1 (install) is timed in the install doc
4. First release	Outlined only

The bigger ideas

Idea	State	What exists now
Workflow-first, agent is one node	Built	Eleven node types; agents are one of them
One workflow, several execution modes	Partly built	condition plus agent workspace checkout plus harness executors. No hosted or container sandbox provider yet
Provider independence	Partly built	Anthropic, OpenAI, Copilot per profile; name: default per server. Not capability-based
Capability-based model routing (model: auto)	Vision	None
Portable agents (one contract reused across workflows and providers)	Partly built	Versioned profiles, and a Library of reusable MCP servers, agents, skills and commands. No standalone agent registry
MCP as a first-class runtime primitive	Partly built	MCP tools through the gateway with ledger and policy; remote Streamable HTTP connections; gateway bridged into harnesses. Health, latency and cost metadata pending
Capability resolution (requires: issue-management)	Vision	Tools are registered by name and revision
Knowledge as a uniform interface	Partly built	Datasets with hybrid retrieval and citations. Git, MCP resources, databases and web are not sources yet
Token optimization as part of execution	Partly built	See Cost and token control
Result caching, savings reporting	Vision	Only measured cost is reported; a savings figure needs a defined baseline per workflow
One graph for chat, automation and coding	Partly built	CLI, API, schedule and web entry points. No web chat or webhook triggers yet
Capability Graph	Vision	The run plan already evaluates requirements against executor capabilities, which is the seed
Workflows exposed as MCP tools	Vision	None

Known limits today

Caution

• Workers are supported on Linux only. The installer is a bash script.
• The macOS install is tested and working, but limits.memory_mb is refused there (no prlimit).
• Native Windows local mode is tested and working, and still labelled experimental. OpenCode steps on native Windows are new and not yet checked by a live run.
• The alpha embedder for knowledge is lexical-hash, not a neural model.
• The Codex executor adapter is experimental and off by default. OpenCode 2 is a preview, off by default.
• The 80/100 confidence bar is enforced per finding but not yet measured against a labelled set.

Next steps toward the vision

1. Capability fields on profiles (capability, strategy) resolved by a simple router, with the choice recorded in the run plan.
2. Tool metadata (permissions, health, latency, cost) in the registry.
3. A savings report with an explicit per-workflow baseline.
4. Web chat and webhook entry points into the same runtime.
5. A container or hosted sandbox provider for coding steps.
6. Capability resolution (requires:), then the Capability Graph on top of the run plan.