From a goal to a shipped pull request

Tharas is not one model in a loop. It is a pipeline of narrow roles with structured hand-offs, physical isolation between tasks, and a gate at the end that opens a pull request on nothing but evidence produced minutes earlier.

  1. 01PlanReads the repo, drafts a spec
  2. 02BuildOne worktree per task
  3. 03VerifyRuns it, does not just read it
  4. 04RepairFixes blockers itself
  5. 05ShipOnly on fresh evidence
  6. 06GovernHalts work that stalls
01

It plans, then builds — nothing to approve#

You give it an objective. It writes a specification and a task list before it writes any code, and then it starts.

  • Reads the repository and queries what it already knows about your codebase
  • States the decisions it made for you and the paths it will touch
  • Nobody is asked to approve it — a service it has no keys for is built against a stand-in, and it says so
02

Each task builds in its own worktree#

Your checkout is never the build surface.

  • Every task gets its own git branch and worktree
  • File tools are jailed to the paths the plan declared — reaching outside fails a gate, it is not a warning
  • Multi-service plans get disjoint directories by construction, so concurrent tasks cannot collide
03

It proves the thing actually runs#

Six deterministic gates, and then the application is really started.

  • Non-empty, in scope, parses, compiles, passes your suite, introduces no new type or lint diagnostics
  • Only diffs clearing all six reach the acceptance review
  • The merged tree is booted on a loopback port and its health check and declared routes are requested
04

When it blocks, it repairs instead of reporting#

A blocked task enters a resolution loop, not a queue.

  • Builds an evidence bundle and searches every knowledge tier at once
  • States a hypothesis before researching rather than after, then ranks repair strategies by confidence, risk and cost
  • Repairs go through the same jailed engineer as everything else — the resolution engine has no file-writing tool of its own
  • A confident success becomes a reusable skill, so the next similar blocker starts from something
05

It ships only what passed — and never waits for you#

Nothing is shipped broken, and nothing waits for anybody to approve it.

  • Goes live only when nothing failed: every task passed QA, any suite it has passed, and the application started and answered
  • Plus no parts contradicting each other and no merge conflicts
  • Work that fails is fixed by the next round — or left out, so your project stays as it was
06

It watches what it spends, and stops itself#

Nobody picks a model, and nothing runs on after it has stopped earning its cost.

  • Routing chooses per call from capability, live price, provider health and measured outcomes
  • It may never hand hard or retried work to a weaker model — asserted in code, not merely intended
  • A per-task governor halts a run that has stopped producing progress. Not a spending cap at an arbitrary line

Where you watch all of this

A terminal, your workspace in a browser, a desktop app, and your phone. One engine behind all four — and one bundle behind the three you look at, so an affordance cannot ship to the web and be missing on a phone. Where it can, it is treated as unfinished rather than as a shortcut.

Where you work → All features →

Common questions

Does Tharas need access to my code?

It runs on your machine, against the repositories already there. What leaves the machine is an entitlement check, the model calls themselves, and telemetry that has no free-form string field.

What stops it shipping something broken?

The ship gate requires evidence produced minutes earlier: every task passed QA, the merged tree’s own suite ran and passed, and the application started and answered its routes. There is no setting that ships unverified work.

What happens when it gets stuck?

A blocked task is diagnosed, searched against what Tharas has learned, given a hypothesis, researched and repaired through ranked strategies. You are asked only once that budget is spent, and you get a report of everything tried.

How many agents are there?

Six make up the build pipeline and cannot be disabled — Architect, Engineer, QA, Forge, Medic and Curator. Optional roles can be benched, which stops them acting on their own but never blocks a direct instruction.

You have read what it does

Try it on a repository you already have.

Point Tharas at an existing repo, or describe something you want built from nothing. The plan comes back before any code does, and you decide whether it starts.

  • Free to download and keep
  • Runs on your machine
  • No account needed for the file