What's live, what's coming.

Describe the work in plain language — no markdown, no config, no code.

Wright takes a sentence and opens a run; the four-agent pipeline does the rest.

Compose the whole workflow on one canvas — sources, agents, gates, outputs.

The canvas: a real multi-node graph you build and run, not a chat thread.

Four Claude agents, one job each — conducted on a visible canvas, not a black box.

Designer → Optimizer → Scaffolder → Reviewer, each phase openable mid-run.

Every decision leaves a receipt you can audit — full visibility, start to finish.

The live run rail records each phase, the spec, the changes, and the agents' reasoning.

An independent security pass runs on every build, in every autonomy mode.

The Reviewer agent: schema validation, secret/credential flags, hook & MCP review.

Put a person in the loop exactly where you want one — approve, edit, or send back.

The review gate, as it appears in a run; the work waits until a person decides.

Run a workflow fully autonomously, or apply gates wherever you want a person in the loop.

Per project: run end to end on its own, or place gates — at deploy, at every phase, or none. Your call, set once.

Take a task from plain language all the way to a shipped production build.

Autonomous end-to-end delivery for the projects that allow it — generation through to a shipped build.

Agents write to other systems only through an allowlist — never beyond it.

Act nodes are constrained to an explicit allowlist; this holds even with the gate off.

See the real dollar cost of every run — deterministic pricing, not estimates.

The pricing engine captures model usage per run and shows cost on the run page.

Watch your competitors daily and turn every move into a pre-framed, buildable recommendation — the only platform that closes the signal-to-build loop.

Each signal arrives with a suggestedBuild brief Claude wrote alongside it, and a one-click handoff to /build that carries the signalId through for traceability. Signals only recommends — the build still passes the human gate.

Score every development on materiality × relevance × recency, so your team sees only what matters to your market.

A durable daily scan watches each team's own company, competitors, and vertical, then ranks a filterable feed across three severity tiers — each signal with evidence, source URLs, and a confidence chip.

An agent synthesizes the raw firehose into named, rankable signals — never a wall of noise. Mute, snooze, and pause to control fatigue.

Claude turns the raw stream into discrete scored signals; per-signal mute, snooze, and pause keep the feed honest. Refine any recommendation with Wright before you build it.

A curated, citation-grounded record of how your org operates — built so an agent can act on your behalf.

An 11-type taxonomy (SOPs, decisions, systems, tribal knowledge, and more) where every statement carries a checkable citation — a repo path and SHA, a URL and locator, or an interview turn.

Populate it four ways: point at a repo, paste URLs, interview your team, or let the agent draft and you correct it.

A repo runs the agentic Extract/Retrieve/Generate/Validate pipeline; URLs build a company profile; the agent drafts and asks only the genuine gaps. Every generated entry lands DRAFT — a human gate promotes it to CONFIRMED.

Every statement grounds in a checkable citation — and the agent's own guesses are labeled unverified, never dressed up.

Each entry traces to its source or is honestly marked an unverified agent draft. The agent's own KB tools are read-only — get_kb_context, search_kb, get_kb_entry, with no write tool.

The agent reads curated, linked summaries, not a raw dump — no context rot, no lost-in-the-middle.

A projected context file (llms.txt style: a heading, a blockquote summary, linked lists) so the agent reads summaries and follows links instead of swallowing the whole record.

An immutable audit log records every change — who, when, why, from what source.

An append-only revision ledger keeps the full history of every entry; an inline human edit creates a manual source and a new revision, so provenance is never overwritten.

Bring the work to the Software Factory instead of standing up your own.

The client-facing offering, powered by Bridge. Productized delivery, delivered.

Every team's work is isolated to its own workspace — scoped on each query, enforced at the database.

App-layer org scoping runs on every read and write, and Postgres row-level security enforces the same boundary at the database — verified on Neon: an unscoped query sees nothing, a scoped one sees only its own workspace.

Give each person the right access — owner, admin, member, or viewer, scoped to their workspace.

A three-tier identity model: a platform operator above each workspace's owner / admin / member / viewer roles, gating who can create, run, manage people, or only read.

An AI interviewer elicits the tribal knowledge humans forget to write down — using the Critical Decision Method.

A cold-start interview built on the Critical Decision Method, seven question types, and 14 mistake criteria — capturing the undocumented know-how, then landing it as DRAFT for a human to confirm.

Enterprise workspaces can run on dedicated, physically separate infrastructure — for data residency and the strictest separation.

A dedicated database per enterprise tenant: the same product on isolated infrastructure. Design complete; provisioning in progress.