Skip to content

TERX v0.5.0 — Save approved workflows as tools

Choose a tag to compare

View all tags
@ixchio ixchio released this 30 Sep 20:12
· 4 commits to main since this release
v0.5.0
42e2d16

TERX v0.5.0 release

Released 2026-10-01.

TERX v0.5 turns an already approved semantic browser workflow into a local,
named MCP tool. This is not a generic agent builder: every call either finds a
matching policy-bound replay, rechecks its evidence, and returns declared live
page data—or refuses.

What shipped

  • browser_tool_save, browser_tool_list, browser_tool_run, and
    browser_tool_delete for local MCP tool manifests.
  • Exact string input schemas and fresh output fields limited to visible CSS
    selector text, page title, page URL, and non-sensitive echoed inputs.
  • Per-run variable values and optional fresh result readers in TerxWorkflow.
  • SQLite schema v4 manifest storage. It retains a scope digest and input names,
    never raw scope identifiers, input values, or result payloads.
  • A real headless-Chrome local golden path for a saved
    check_order_status(account_id, order_id) tool. It reconnects the browser
    CDP session, changes both inputs, returns the new status, exercises a slow
    page load, and refuses an ambiguous target.
  • A reproducible 20-second demo built
    from the unaltered frames of that local golden path. Run
    python3 examples/capture_saved_tool_demo.py --output /tmp/terx-demo-frames
    followed by bash docs/assets/render-v0.5-demo.sh /tmp/terx-demo-frames to
    reproduce it.

Cost and benchmark boundary

ReplayCostLedger no longer infers LLM savings from action count. A TERX replay
reports the model calls made by TERX itself (0); provider tokens, calling
assistant cost, and browser cost require host usage records. The included local
evaluation is a fixture proof, not a Browser Use, Stagehand, or production-site
comparison. See benchmark protocol.

One separate Browser Use Cloud v4 smoke run was measured using gpt-6-luna,
low reasoning, a US browser, and a $0.25 cap. It returned the expected final
host (www.iana.org) in 16.017 seconds, consuming 48,492 input tokens and 218
output tokens for $0.002516. It is a real provider record, but a different
public task—not a TERX-versus-Browser-Use performance claim.

The release notes also include a separate native Browser Use rerun_history
smoke record: after a Chrome restart it completed its stored public-page actions
in 4.032 seconds, but consumed 1,033 Groq tokens for the rerun summary. TERX
does not use that fact to claim superiority; it documents the architectural
difference between a framework rerun and TERX's constrained replay layer.

Deliberate limits

  • Chromium/CDP only; no cross-browser or cloud browser fleet.
  • No proxy/stealth/CAPTCHA stack, credential vault, visual builder, or hosted
    operations layer.
  • Saved tools share the local TERX cache/server. They are not a hosted registry.
  • Browser Use remains experimental: normal Browser Use action traffic is not
    safely captured by TERX's semantic recorder yet.
  • There is no automatic replay self-healing. Refusal is the intended safety
    outcome when the contract no longer matches.

Verification performed

ruff check .
pytest tests/ -q
python3 -m terx.evals.local_suite
python3 -m build
python3 -m twine check dist/*

The release workflow reruns this suite under Python 3.11, 3.12, and 3.13.

Assets 5
Loading