Flow canvas
Groups, nodes, typed connections, variables and presentation metadata. The accessible outline projects the same FlowSpec, so the script is readable without a mouse.
Runs on your machine.
Product
One place for the branch, the handover, and the reason it answered that way.
A bot that is edited in one tool, watched in a second and explained in a third has three versions of the truth. Chawo keeps the script, the run and the revision in one register.
The canvas holds the script, and the outline holds the same script.
Groups, nodes, typed connections, variables and presentation metadata sit on a spatial canvas. The accessible outline is not an export of it — both are projections of one FlowSpec, and neither owns state the other cannot see. That is what makes the script reviewable by someone who did not draw it.
The handover is a node, not an escape hatch.
Where the bot stops and a person continues is drawn in the same script as every automatic branch. The run records the handover alongside the path, the outcomes and the repairs, so the question of what the bot had already promised before a person arrived has an answer in the ledger rather than in someone's memory.
One run, read in order.
A run is one execution attempt against one revision. It carries the path taken, the outcomes, the timing, the repair steps, the handover and the delivery facts. Where a value was never established it stays unknown: a missing fact is not a negative result, and it never becomes a success because the interface had nowhere else to put it.
Analytics across runs, not a dashboard about them.
Behaviour aggregated over many runs, so a branch that quietly started failing last week is something you notice. The analytics surface is in memory today and reseeds when the process restarts, which makes it a working instrument rather than a system of record.
Revisions are immutable, and templates do not mutate an agent.
A validated draft compiles into an immutable revision carrying receipts. Templates are versioned packages that import as a new draft, so adopting one never rewrites an agent that is already answering. The facts you collected keep pointing at the logic that produced them, including after you changed your mind.
What is in the box today.
Run ledger
Path, outcomes, timing, repair, handover and delivery facts for one execution attempt against one revision.
Runs on your machine, over redacted or synthetic data.
Analytics
Behaviour aggregated across runs, so a pattern is something you notice rather than something a customer tells you about.
Runs on your machine, in memory, reseeded on restart.
Revisions and templates
Immutable revisions with receipts, and versioned templates that import as a new draft instead of quietly mutating an agent that is already answering.
Runs on your machine.
Local-first is a position here, not a stage on the way somewhere.
- The workbench runs on your machine and does not contact a live runtime.
- It reads redacted or synthetic projections, never a production database.
- It refuses any request that does not arrive over loopback.
What Chawo does not do.
Refusing work is part of the work. These are the boundaries as of this build, and they are on this page rather than in a footnote.
- Telegram delivery is at launch
- Today Chawo does not deliver messages to Telegram. The Chawo runtime sends them at launch.
- Model calls are at launch
- Today the workbench records provider and prompt identities as non-secret facts. The Chawo runtime calls the model at launch.
- It does not write to a CRM or an ad platform
- There is no integration that writes outward, and none is presented as pending configuration.
- It does not read a production database
- It reads checked-in redacted or synthetic projections. That is a design decision, described on the security page.
- It is not hosted, and there is no sign-up
- The current build runs on one machine and answers only over loopback. Multi-tenancy and authentication are not built yet, which is why early access is a list rather than a form that creates an account.