Keydler
Keydler is local-first, human-supervised task memory for people working with AI agents. It
keeps completed work, active rules, and rejected approaches on one page, so an agent that
loses context in a new conversation does not repeat work it already finished or retry an
approach that was already ruled out.
The problem
When an AI agent loses its context, a conversation summary keeps the highlights but drops
the constraints: the approach already ruled out, the dependency already refused, the step
already completed. Keydler keeps that memory on the page itself, in the browser's local
storage, and exposes it to AI agents through WebMCP
(document.modelContext.registerTool) so an agent can read and write the same
state a human is looking at.
The thirteen WebMCP tools this page registers
Four read tools and nine write tools. Writes require an active task and are versioned, so
a write based on stale data is refused rather than silently applied.
-
resume_task — the canonical task state: id, version, rules, rejections,
next step.
-
what_changed — what was written since the version the caller holds.
-
read_task_detail — paginated detail: full evidence, full reasons, older
work, credential names.
-
search_task — search one term across steps, evidence, rules and
refusals.
- log_step — record a completed step and its evidence.
-
attach_evidence — attach proof that arrived after a step was logged.
-
set_next_action — redirect the task without inventing a step to hang it
on.
- add_constraint — propose a rule to follow.
-
reject_approach — propose ruling out an approach, with a mandatory
reason.
-
add_decision — record the reasoning behind a choice, before a summary
loses it.
- ask_human — record a blocking question instead of guessing.
-
request_approval — ask permission for something irreversible, and wait
for a human decision.
- complete_task — close the task with a hand-over summary.
Keydler runs entirely in the browser: data is stored locally in IndexedDB, with no account
and no server required to use it. The source is open under the MIT license at
github.com/kymrapro-del/ChatGPT-WebMCP.
This page needs JavaScript enabled to open or create a task.