Text Brief
Turn supplied source material into a faithful, reviewable brief—with fully local inference.
RESULTBoard briefExport
Claims18Questions2Coverage96%
FaithfulStructuredOffline
The product layer for agent workflows
Turn a proven agent workflow into a Mac app anyone can use.
Bring the workflow that already works in Pi, Hermes, OpenClaw, or your own stack. The factory adds a focused interface, pinned tools, an isolated runtime, managed inference, narrow permissions, evaluations, signing, and updates.
macOS firstlocal inference todaycurrent prototype
text-briefapplication source
VALIDbrief.skill.mdportable reasoning
12 KBinterface.yamlinputs · plan · result
policy.yamlfiles only · network denied
evals/30 representative runs
RUN TB-04802:18
Creating brief
74%Validate the claims.
✓
Map the source18 claims · 6 supporting passages
Done✓
Resolve contradictions2 tensions retained as open questions
Done3
Validate the claimsChecking every section against source text
Running4
Build the briefSummary · evidence · decisions · gaps
NextBENCHMARK30 / 30runs passed
Permission policySelected files only
SKILLINTERFACETOOLSPOLICYEVALSSIGNED APP
Apps, not agent demos
Not every useful agent should become a chatbot.
The strongest candidates are narrow, repeatable jobs with a clear start, governed actions, and a finished result. Two apps already exercise the shared factory; four show where it can go next.
File Inventory
Inventory only selected files with names, sizes, and checksums—without packaging a model.
4 FILES SELECTEDSHA-256
PDF
research.pdf2.8 MB
7f3a…91b2MD
notes.md18 KB
01c9…44a1CSV
claims.csv42 KB
a923…e101Explicit selection onlyExport inventory
Research Dossier
Turn a bounded set of sources into a decision-ready dossier with evidence, disagreements, and open gaps intact.
PDFURLDOCCSV
→DECISION DOSSIEREnter the market?RecommendationEvidence mapRisks & unknowns
Bookkeeping Pre-close
Match receipts against a ledger, surface missing evidence, and propose reviewable changes—without posting them.
APRIL PRE-CLOSE7 items need review
✓
OpenAIReceipt matched · $200.00
Ready!
FigmaPayment found · receipt missing
Review?
Unknown chargeNeeds owner classification
AskRelease Review
Inspect a packaged build, run declared smoke tests, and produce a release decision with exact evidence.
R
ARTIFACTRelease.appv1.4.0-rc2
✓ Signature valid✓ Offline launch✓ Update apply○ Rollback test
DECISIONHold for rollback result
Voice Review
Apply a house style across a batch of drafts, explain each edit, and leave the final decision to the author.
BEFORE
7 edits We leverage a comprehensive solution to enable…
AFTER
Use one focused workflow to turn the source into…
Built now means the app exercises the current shared factory. Designed for is an example of the product shape, not a shipping claim.
The market gap is capability delivery
Codex provides the workshop. The factory ships the appliance.
A technical user can often reproduce the underlying behavior. The hard part is making it dependable, legible, governed, and distributable to everyone else.
A pile of capability.
$ git clone workflow$ brew install python node llama.cpp$ export PROVIDER_KEY=••••••$ agent --skills ./skills --tools allError: model server unavailable- Blank chat and implicit instructions
- Broad credentials and permissions
- Dependencies the recipient now owns
- No release, rollback, or behavioral contract
COMPILE
A finished product.
A
AGENT APPLICATIONText Brief.appSigned · self-contained · updateable
- One normal installation
- Purpose-built inputs, progress, and result
- Declared authority and approval moments
- Evaluated behavior, cancellation, and cleanup
A reusable application compiler
The reusable part is everything around the skill.
A new product should mostly require a portable skill, manifest, interface schema, deterministic tools, fixtures, and branding—not a new runtime, policy system, updater, or release pipeline.
01APPLICATION SOURCE
Portable skillReasoning, sequencing, validation, escalation
Workflow + UI schemaInputs, plans, approvals, results, exports
Deterministic toolsRegistered scripts, APIs, storage, OS actions
Capability policyMaximum files, domains, secrets, and actions
Evaluation casesRepresentative, adversarial, cancellation, cleanup
02SHARED FACTORY
One product runtime
Electron shellAgent contractModel managerPolicy brokerPrivate stateLifecycle controlUpdaterDiagnostics
REUSE TESTMaterially different applications ship without modifying the core runtime.
03FINISHED ARTIFACT
MAC APPLICATION
Focused App.app
One install. One job. Explicit authority.
Intended toolchain
Five commands from source to controlled release.
Each command makes a different promise explicit: valid authority, measured behavior, reproducible packaging, exact-artifact smoke tests, and a governed release.
➜skillapp validate text-brief
✓ manifest schema
✓ registered tools
✓ capability policy
✓ offline contract
VALID · 0 unsupported capabilities
Task-specific product experience
Not a blank chat box.
The interface carries the workflow: structured input, a reviewable plan, bounded approvals, visible progress, safe cancellation, and an inspectable result.
INPUT CONTRACT
The user supplies the source, purpose, and constraints through fields the app can validate before any model runs.
Create a faithful brief
Add the source and tell the app what the brief must preserve.
Keep exact claimsSeparate facts from inferenceSurface unresolved gaps
Approve the job before it runs.
The app translates the skill into visible, cancellable steps.
1
Map the sourceExtract claims, evidence, decisions, and unresolved questions.
≈ 35 sec2
Resolve contradictionsRetain tensions rather than inventing a single answer.
≈ 20 sec3
Validate every sectionTrace the brief back to supplied source passages.
≈ 55 sec4
Render and exportWrite only to the chosen location.
Approval◇
Authority for this runRead supplied text · bundled tools · no network · ask before export
Validate the claims.
Checking every section against source text.
✓
Source map complete18 claims · 6 evidence clusters · 2 tensions
09:42:11✓
Contradictions retainedNo unsupported reconciliation introduced
09:42:29Validating section 4 of 6Checking “managed inference” against source passages
Running4
Render final briefWaiting for validation pass
QueuedOffline enforced0 network requests · 1 owned process
Review the finished brief.
Every section remains traceable to the source.
BriefClaims 18Questions 2Run log
DECISION BRIEFAgent App Factory
Source coverage 96%Recommendation
Prove that several materially different applications can share one dependable runtime and packaging system before expanding the capability surface.
CORE INSIGHTSkills make expertise executable. The factory makes that expertise installable.
Evidence
Two applications already test both agentic and deterministic paths.
Open gap
Public signing, notarization, update apply, and rollback remain release gates.
FOCUSED UI
The task is encoded in the product—not hidden in a prompt.
LIFECYCLE
Pause, cancel, recover, and clean up are first-class behavior.
Narrow authority by construction
The app can only ask for what its manifest declares.
Every agent action crosses a deny-by-default broker. Capabilities are scoped to one finite run, consequential actions can require one-action approval, and access is revoked when the job ends.
Deny by defaultUnsupported capabilities fail validation instead of receiving broad access.
Selected resources onlyFiles and export locations come from explicit user choice.
Approval at consequenceThe app can plan freely while exact writes and external actions stay gated.
Authority expiresRun-scoped capabilities and owned processes are revoked and cleaned up.
Current proof
One factory. Two materially different apps.
The claim only matters if agentic and deterministic products can share the shell, policy, lifecycle, diagnostics, and packaging system without forcing every task through an LLM.
AGENTIC APPLICATIONText Brief
LOCAL MODELA real semantic workflow through Pi Agent Core, a private bundled llama.cpp runtime, and Qwen3-4B—with no account, API key, subscription, or hosted fallback.
Semantic executionSafe cancellationRuntime-slot reuseExportOffline enforcementProcess cleanup
DETERMINISTIC APPLICATIONFile Inventory
NO MODELA focused utility that inventories explicitly selected files with names, sizes, and checksums while packaging no model and no agent runtime.
Explicit selectionDeterministic toolsPrivate app stateCSV exportSame policy brokerSame packaging path
PACKAGED TEXT BRIEF BENCHMARK
Every run passed.
Structured output, tool behavior, cancellation, and semantic quality all passed across the 30-run benchmark.
WHAT THIS PROVES
The product layer can be shared.
- ✓ Real semantic execution in the packaged path
- ✓ Deterministic apps are not forced through an LLM
- ✓ Cancellation and process cleanup are testable contracts
- ✓ Offline behavior can be enforced, not merely promised
Current prototype: evidence for the architecture, not a claim that public distribution is complete.
Who the factory is for
Bring the expertise. Keep the product layer.
Workflow builders
You already have a skill or agent workflow that works. Ship it without rebuilding model management, policy, UI, packaging, evaluation, and updates.
Domain experts
Your method combines judgment with repeatable machinery. Turn that expertise into a focused product instead of a prompt library or consulting-only workflow.
Controlled teams
You need reproducible behavior, declared authority, review points, and evidence—not a general agent with ambient access to company systems.
Build the workflow once
Skills make expertise executable. The factory makes that expertise installable.
Prove the shared factorySeveral useful Mac apps · local and deterministic paths · dependable runtime
Expand explicit execution modesExisting AI subscriptions · paid APIs · more local models
Add new product primitivesScheduling · event triggers · supervised computer use · catalog distribution
Early access
Your workflow already works.
Make it installable.
Tell us what you have built—or what finished agent app you wish existed. The earliest design partners will shape the first reusable application contracts.