install the engine, add a kit, keep the receipts
Your coding agent says the tests pass. Flow Agents installs into that agent and makes it show the work before it can call the job done — then writes a file recording what ran, what passed, and what it skipped. Anyone can re-check that file without trusting us or the agent. This page is the map: install it below, then open whichever part you want to look inside.
Start here
Two minutes: put the engine in a workspace you already use, then take one of our published receipts and recompute under two validators yourself. Both commands on the right are the ones this site's own CI runs on every change.
$ npx @kontourai/flow-agents init
# asks which agent you use, then wires it up:
# Claude Code, Codex, Kiro, opencode, and pi —
# gates block on Claude Code and Codex; the rest
# are advisory out of the box (matrix on /trust)
Know what it costs you before you paste it. It installs into the directory you run it in — 18 top-level entries, including a README.md and a CLAUDE.md. If your agent already has Flow Agents hooks at the user level, it warns you about the collision rather than quietly running everything twice. Run it at the root of a repo you can commit, and read the diff before you push. Nothing leaves your machine unless you point it at a Console.
# download a bundle from /receipts/, then read it
# with our own validator
$ npx @kontourai/[email protected] report \
--input flow-agents-delivery.trust.bundle --format summary
Claims: 34 (proposed: 1, assumed: 1, verified: 32)
Stale: none
Disputed: none
Transparency gaps: 1
# (the report names that gap in full — trimmed here)
# now run the same file through a different
# implementation, one that reads the open spec's
# schemas instead. ajv is hachure's schema engine,
# not a second opinion; both are pinned to the
# versions CI uses.
$ npx -y -p [email protected] -p [email protected] \
hachure validate flow-agents-delivery.trust.bundle
valid TrustBundle (schemaVersion 5)
Two implementations, one file, same verdict. Surface is ours; the hachure CLI is written against the open format instead. Honest scope: both are still Kontour-maintained, so that is implementation independence, not organizational independence.
We run this on ourselves. This site's own pull requests can't merge until a required Trust Verify check passes, and the bundle it produces is published on the receipts page.
The kits
A kit is a workflow the engine runs: what the steps are, what evidence each one has to produce, and when it stops to ask you. How kits work →
Under the hood
You just validated a receipt. Everything below is what had to happen for that file to exist, starting with the format — which isn't ours. Take any one layer on its own, or let init assemble the whole chain and never look at this again.
Every piece above is on npm today. Two things are preview-grade and say so in their own docs: the AWS Strands framework adapters, and conformance certification for third-party kits. Lane one is the odd one out — the receipt format is published outside Kontour, so we can't quietly change what a receipt is allowed to say.
Built on it
It sits on that chain rather than being a link in it, which is why it isn't a lane in the map. Start here when you want the complete workflow instead of the parts.
Turn messy source text into reviewed, source-linked data you can recheck and export.
The finished app: install it and start reviewing. Survey is the review machinery inside it, if you would rather build your own screen around it.
@kontourai/fieldwork · v0.6.0
The territory
Every product is named for the surveyor's craft, because that is the job: go into the field, measure carefully, and bring back records you can defend. Regions are problem domains; Surface sits at the summit because it is the reference everything measures against. A line is drawn only where the relationship exists in code — gold for the grounded-knowledge pipeline, blue for artifact flows, dotted where a product speaks Hachure, the open trust format. Hover any marker for its role, what it produces, and what consumes it.
buildSurveyTrustBundle mints the TrustBundle on behalf of its consumers.producesTrustBundle → Surface shapeusessurfacekind: trust.bundle: validated, report-derived, freshness re-checked on every evaluation.producesrun projections → consoleusessurfacekit.json contract.built-inbuilder (the delivery workflow) · knowledge (gated knowledge store) · release-evidence (agentless CI gate)externalgovernance ships with veritas and gates the bundle veritas readiness emitskits: builder · knowledge · release-evidence · governance
More from the lab
Smaller pieces we built for ourselves and published anyway. Each does one thing, each works without the rest, and every one of them is on npm today — nothing on this list is a name you can't install.
@kontourai/cli
The kontour command. It works out which Kontour packages you already have installed and hands your command to the right one. It never reaches for the registry, downloads a missing package, or changes a version behind you — so the same command does the same thing offline.
@kontourai/datum
Your model registry. Define providers, models, and roles once, then resolve which backend, which model, whose key, what URL. It doesn't call models for you — apart from four opt-in commands that exist to test a connection.
@kontourai/bearing
Which model is actually good at what, and where that judgement came from. Compiles observations into catalog snapshots you can diff, and ranks models for a given runtime — without calling any of them.
@kontourai/relay
One way to call a model whether it sits behind a vendor SDK, a local engine, a hosted service, or a command-line harness, so changing providers doesn't mean rewriting the call. Each adapter has to declare what it can't do — streaming, tool calls, token counts — rather than quietly approximating it.
@kontourai/dispatch
Chooses which model handles a job from rules you write — required capabilities, a token and cost budget, retries, fallbacks — and hands back a record of which one ran, why it was picked, and what it actually spent. Budgets can be held across concurrent calls and across restarts.
@kontourai/conduit
Installs an agent's skills, hooks, commands, and context files into whichever coding tool someone is running, through one adapter instead of one per host. It probes each host for what it genuinely supports, and the install receipt lists every file written with its checksum — no file contents, no credentials.
@kontourai/traverse
Pulls structured fields out of documents against a schema you supply, and hands back proposals for a person to review — each one still carrying the passage it came from.
@kontourai/forage
Fetches untrusted URLs safely and keeps a snapshot of what came back, so a crawl can be replayed instead of re-guessed. Refuses requests aimed at your internal network.
@kontourai/lookout
Watches sources you re-check over time and tells you what changed since the last look. Built on Forage's snapshots; it doesn't extract fields itself.
@kontourai/plumb
Runs your deployment's own health check on a schedule, and when it fails, hands the failure to an AI agent working in a separate clean clone that opens a pull request or an issue. The live checkout and production data stay outside anything that agent can reach.
@kontourai/ui
The interface pieces our own screens are built from: buttons, panels, badges, dialogs, form controls, and the CSS variables behind them, for React or plain custom elements. Useful if you are building an operator screen that should match ours; skip it if you already have a design system you like.