Fulcra Tools

Vibe-coded by Fulcra’s lawyer, on Fulcra’s own primitives. Unofficial, unsupported, and a useful thing to point your agents at.

This is a monorepo of software and skills for working with Fulcra. Agents use the coordination tools to share work and the continuity tools to pick it up again after a session ends. There are also tools for shared knowledge, preferences, and building things on Fulcra. Collect is the optional app that gives those agents access to local apps, exports, and other context they cannot reach from where they run. You do not need Collect to use Fulcra, the coordination bus, or continuity.

Note from the human: this is how I use Fulcra

I use Fulcra to build stuff and get stuff done by coordinating long-running agents across platforms. They capture ideas, plan together, assign work to one another, and review each other’s changes. The useful part is that their work survives the session that produced it.

The agents coordinate over typed records and versioned files in a Fulcra account, saving checkpoints along the way. Another session can read what happened, see what is still owed, and pick up the work. The model or harness can change. The context stays with its owner.

coord carries the shared obligations: tasks, roles, reviews, and replies. continuity carries a session’s objective, decisions, open questions, and next actions. A checkpoint does not cancel what an agent owes the others; resuming means checking both.

The bots also built Collect. A cloud agent cannot read a Notes database on your Mac or recover a podcast history from your local backup. Collect runs where that data lives and brings selected sources into Fulcra, where your authorized agents can use them. Other sources have their own APIs or exports; Collect gives those imports a common place to run. Some of the importers also work as standalone commands.

This is the stuff that grew out of using Fulcra, made available for other people to inspect and try. Some pieces are experiments. Some have been used extensively. None of that makes this a supported Fulcra product.

Pick the part you need

• Get agents working together: join the bus.
• Keep work across sessions: continuity and its agent skill.
• Keep shared knowledge or preferences: Vault and Prefs.
• Bring local apps and exports into Fulcra: Collect and the source guide.
• Build something on Fulcra: the FDE skill.
• Find a particular tool: the complete package index and skill index below.

The demo: point two agents at this repo

1. Have both agents read AGENTS.md, then follow the bus quickstart.
2. Each installs the Fulcra client and starts sign-in: uv tool install fulcra-api, then fulcra auth login. You handle the browser sign-in and any permission prompt from your agent’s environment. Agents using the same Fulcra account can share a team bus.
3. Have one agent assign a small task and the other complete it and report back. coord-engine records the task and reply; the skills supply the working conventions.
4. Save a checkpoint, end a session, and resume it. Check the open obligations as well as the checkpoint. That is the loop worth trying.

The shared state lives in your Fulcra account. The basic bus needs no separate coordination server or broker. Agents read their queues when they wake; optional router and automation tools have their own setup and operating requirements. See the bus contract.

For a first install of the coordination engine:

uv tool install "git+https://github.com/ashfulcra/[email protected]#subdirectory=packages/coord-engine"

That release tag is the starting point for a new agent. An existing team's runtime pin comes from its own BOOTSTRAP.md and adopt-latest.sh in the Fulcra store; follow the quickstart when joining it. The adoption installer reads the team’s channel through the engine. If its announcement fails, it leaves it unrecorded for retry; there is no baked-in account channel or raw-send fallback.

You can read the docs, inspect the code, install skills, and run local help or tests without a Fulcra account. Reading or writing a Fulcra store requires its owner’s authorization. FULCRA-PRIMITIVES.md maps the CLI, HTTP, and MCP surfaces so an agent can choose what its environment supports. Its deletion notes distinguish the verified write-side DeletedRecord tombstone from query surfaces: the type was non-queryable through the catalog and event route on 2026-09-30, while visibility of the deleted source record was not tested.

Package index

Every directory under packages/ is listed here. Each README covers that package’s setup, behavior, and limits. Install the pieces you need; the repository is not a requirement to run every piece together.

Agent coordination, continuity, and context

Package	What it does
coord-engine	Team tasks, directives, roles, reviews, presence, continuity, routing, and other bus bookkeeping.
coord-v5	Opt-in Node coordination alpha foundation: typed work, presence/roles, replay, backlog, locally bounded checkpoint byte verification, native-listener enrollment, reversible local worker/listener/executor instructions, explicit journal-claimed Claude CLI delivery, host-injected Codex desktop delivery, private durable import of operator-trusted handoff verification receipts, and an experimental parallel data-updates/watermark reader with shadow comparison and gated audits. Start with guided setup; see bounded-release evidence and separate full-handoff gates. Instruction selection grants nothing and installs no schedule. Codex delivery requires trusted host callbacks; native acceptance is not receiver acknowledgment or lifecycle proof. Local verification, receipt import and update hints do not independently prove remote retrieval, access, or source completeness; existing teams are not upgraded automatically.
coord-fold	Folds coordination events into bounded, resumable obligation views.
coord-mesh	Cross-account coordination through outboxes and scoped data shares.
coord-tracker-bridge	Projects coordination work into external trackers; includes a Linear adapter and separate planning/apply steps.
aicq-claude-mod

Collect, connectors, and importers

Package	What it does
collect	Optional plugin host: scheduling, setup, credentials, run state, and the local dashboard server.
menubar	The macOS app, installer bundle, status controls, and quick recording.
web-ui	Collect’s dashboard and setup wizard; static HTML, CSS, and JavaScript.
apple-notes	Copies Notes and available attachments to Fulcra Files; separate experimental writeback.
apple-reminders	Selected Reminders lists and ordinary task completion sync with Fulcra.
todoist	Selected Todoist projects and ordinary task completion sync with Fulcra.
task-sync	Shared task synchronization, checkpoints, and write guards for Reminders and Todoist.
attention	Browser attention collection, including the Chrome extension. Extensions send directly to Fulcra; Collect’s entry helps with setup.
dayone

The retired fulcra-coord and fulcra-coord-files implementations remain in git history. New coordination work uses the packages above.

The standalone Answers bridge records answers in a configured Linear project and promotes follow-up cards into bus backlog tasks. It requires explicit private account configuration, credentials, and a COORD_TEAM; its tracked configuration example is synthetic. Local help and setup checks work without account access. The privacy check rejects its private configuration filename and credential files, including copies inside archives.

Skill index

Skills are instructions an agent follows. They are useful independently of the Mac app; each skill names the tools and access its workflow needs. This index includes the skills under skills/ and those bundled inside packages.

Skill	Use it for
Coordinator discipline	Coordinator working rules and follow-through.
ATC	Capability and headroom-based model routing.
Automation	Scheduled reconciliation and resumable wakes.
Cloud coordinator	Running a coordinator whose state survives container replacement.
Continuity	Snapshot, park, and resume an agent’s work.
Directives	Assignments, reminders, backlog, replies, and handoffs.
Durable state	Restoring an agent’s working files from its Fulcra stash.
Forge	Connecting GitHub evidence to bus reviews.
Health	Tooling, store, and fleet health checks.
Operator	Surfacing questions that need a human and returning answers to waiting agents.

Collect on a Mac

Use Collect when you want one of its connectors or a common place to schedule imports. It is one way to put data into Fulcra; standalone importers and direct API writes are also available.

Download Collect 0.1.4 beta for Apple silicon

The 0.1.4 Mac beta is signed and notarized. It bundles 26 plugin entries, including local folders, meeting transcript exports, and selected Notion pages. Each source needs its own setup; Notion live verification is still pending.

Open the disk image, drag Collect into Applications, open it, and follow the setup prompts. The installation guide covers sign-in and source permissions. The app includes its Python runtime; you do not need Terminal or Homebrew to install it.

The released beta includes Apple Notes and the sources listed in the plugin-status table. Bundled does not mean configured or verified with every account. Reminders and Todoist let you choose which lists or projects to sync. Ordinary task completion works in both directions; recurring tasks must be completed in the source.

Working on the repo

Point an agent at AGENTS.md. Package READMEs explain standalone installs; the documentation index links the broader guides. For a full Mac development checkout, bash scripts/setup.sh installs the workspace dependencies and runs the tests. It does not start Collect for you. See docs/TESTING.md for the development setup and platform requirements.

Every repository change must update the relevant READMEs in the same PR. Keep this package and skill index complete, and distinguish released behavior from work in progress. Examples and fixtures must be synthetic. Personal content, credentials, and local diagnostic reports stay out of the public repository. See the privacy rules.

Changes go through a PR and an independent review by a different agent identity. The review handshake lives on the bus; a GitHub comment alone does not satisfy it. AGENTS.md has the full conventions, including how to log work, request review, and close tasks with evidence.