Anna

An AI companion for personal tasks, business workflows, and project collaboration.

Anna is a personal open-source project exploring how an AI agent can carry work from a conversation through to a reviewable result. The local-first desktop app brings tasks, connected business systems, and collaboration together in Home, Cowork, and Crew.

General Agent + Jev Crew · Experimental Source Preview · macOS arm64 · MIT License · CI

中文 · Product overview · Quick start · Current status · Codex pet · Development diary

Product overview

One Anna, three workspaces. Start with your own task, work with a connected business system, or coordinate a project with people and Workers.

Home · Start a task or create something reusable

Home is Anna’s starting point: Chat and Create sit together with common task starters, conversation history, workspace selection and Goal mode. Ask a question, bring local files into a task, or start creating a reusable resource.

Cowork · Bring connected business data into the workspace

Cowork puts a connected system’s data and workflows alongside Anna. The Hiker example shows customer and contract metrics, collection progress, aging and a follow-up entry; the sidebar also exposes the reimbursement assistant. Available operations depend on the connector and its permissions.

This screenshot uses the repository’s synthetic Hiker sample data served through a local MCP fixture. It illustrates the integration and interface; it is not a production Hiker tenant or evidence of business writes.

Crew · Coordinate people, Workers and project work

Crew brings the project list, team and SOP templates together with the task graph and project channel. See dependencies and parallel work, follow assignments, and keep discussion, deliverables and review decisions in context.

All three overview images show the current desktop app with full navigation, captured on October 9, 2026. They use isolated demo data; the Crew project starts from the built-in example.

Home · From a local file to an analysis report

In this live demo, Anna read a synthetic CSV, calculated conversion and return on spend with Python in the sandbox, and saved a report. This close-up shows the resulting table and chart inside the conversation.

Crew · Read the artifact before deciding

The full-width reader keeps the artifact version, source and project channel together. Approve the result or return it with comments. This view uses the built-in example.

Jev · Inspect the suggestion, then assign

A live Jev call suggested the writing Worker for the demo’s content-editor task. The picker shows the suggestion’s source and keeps Accept assignment separate from generating it; manual selection remains available.

The displayed Jev timing is one request, not a new benchmark. Screenshot provenance.

Quick start

Requirements: macOS arm64, Node.js ≥22.19.0, Python 3.12, and uv.

git clone https://github.com/Foxtailsss-Andy/Anna-Agent.git
cd Anna-Agent
npm ci
uv sync --locked --extra dev
ANNA_OMP_BUN_ARCHIVE_URL=https://github.com/oven-sh/bun/releases/download/bun-v1.3.14/bun-darwin-aarch64.zip npm run harness:omp:prepare
npm run desktop:run

To run Agent tasks, configure your model provider locally; business features also require their connector configuration. The current documented model setup uses DeepSeek through an OpenAI-compatible transport.

Follow DEVELOPMENT.md for configuration paths, state isolation, and troubleshooting. Keep model credentials, connector secrets, and runtime state outside the Agent-readable task workspace. External model providers and connectors receive the requests you configure; local-first describes where the desktop app and its state run.

This update

• Read and act on files: select a workdir to inspect it; explicitly enable write permission to edit files and run commands in the macOS process sandbox. Configure MCP tools through protected Host settings.
• Continue a bounded goal: keep a plan across Runs, pause, resume or stop, then review the result. Malformed tool-argument JSON produces recoverable failed-tool feedback instead of aborting the whole Run.
• Keep collaboration usable: Crew retains channel and Anna history, offers confirmed graph/assignment proposals, readable tables/charts, streaming activity and Trace. Queued steering is distinguished from a message the model has consumed.

Jev remains a bounded assignee judgment. The historical comparison below reports 85.23% lower p50, 94.71% lower off-peak estimated cost and equal 22/22 label agreement. Total reported tokens increased 57.98%; this is not a whole-Agent-task speedup claim.

Ask for an assignee suggestion, then decide

Crew can suggest a person or Worker for one unassigned task. Open the existing member picker, request a suggestion, inspect its source, and explicitly accept it or choose manually. A unique exact-role match uses a rule; Jev handles other eligible semantic choices and can abstain. Suggesting someone does not assign or run the task. This source preview adds explicit suggestions; see the changes, evidence, and limits. The RC2 source release remains a separate release record.

Measured Jev comparison

On 24 synthetic heldout cases, 22 were eligible for model judgment; two hard-precheck cases made no C/D call. Both models matched the frozen labels on 22/22: 12 correct recommendations and 10 correct abstentions, with zero wrong labels, API or format errors. Existing role rules matched 12/22 on the same subset.

Jev had 85.23% lower p50, 84.33% lower p95 and 97.35% lower reference estimated cost, with equal label agreement. Reported total tokens increased 57.98%; different tokenizers prevent an equivalent-compute interpretation. The cost figure uses DeepSeek peak/cache-miss prices for budget accounting, not a bill. At this run's off-peak rates with the same cache-miss assumption, estimated cost is 94.71% lower; cache hits may reduce DeepSeek cost further.

This is one small sequential comparison of model judgments, not production accuracy, browser end-to-end latency or measured speedup of the previous complete Crew/Host/OMP path (B). The 8 development cases are reported separately. See full results, denominators and method.

How work moves through Anna

Anna advances a task through a decide → act → observe → decide again loop. The Harness surrounds that loop with context, permissions, execution limits, persistent state, and result checks.

flowchart TD
    Task[Task from Home / Cowork / Crew] --> Context[Load context and authorized Memory]
    Context --> Decide{Decide the next step}
    Decide -->|Action needed| Gate[Check permissions / request approval if needed]
    Gate -->|Allowed| Act[Execute a tool or call a connected system]
    Act --> Observe[Read the result / record success or failure]
    Observe --> Decide
    Decide -->|Ready to finish| Check[Validate execution records and end state]
    Check --> Outcome[Record the outcome and available artifacts]