Accension — Your local-first AI engineering Workbench

Turn software requirements into structured plans, assign work to AI agents, and track execution through testing and review.

Accension is open source, self-hosted and local-first. No hosted Accension account is required. Use local models or optional cloud providers; authorized cloud inference can send project context externally.

Version 1.1.0 adds Workbench: projects, product requirements documents (PRDs), dependent tasks, approved execution and completion review. Routing and Workbench are two views of the same application. Release notes · Upgrade notes.

Get started · Workbench guide · Documentation · Releases

The problem

AI can write code, but delivering a feature also needs a plan, dependencies, permissions, tests and review. Those decisions often get scattered across chats, terminals and project boards. Accension connects them and keeps evidence of what actually ran.

Related MCP server: engineering-mcp

See Accension in action

From requirement to verified work

A PRD states the problem, scope and acceptance criteria. A Work Item turns part of it into a bounded task. Dependencies determine when it can proceed. An agent performs execution; routing chooses eligible models for its roles.

flowchart TB
    subgraph Plan[Plan the work]
        direction LR
        Requirement --> PRD[PRD approval] --> Items[Work Item import]
    end
    subgraph Authorize[Prepare and authorize]
        direction LR
        Dependencies --> Agent[Agent assignment] --> Approval[Execution approval]
    end
    subgraph Verify[Execute and inspect]
        direction LR
        Execution --> Validation[Independent validation] --> Review[Completion review] --> Done[Done]
    end
    Items --> Dependencies
    Approval --> Execution

Watch the complete 51-second demo · 8-second preview · Gallery and capture provenance

Real browser recording of an isolated greeting project. Deterministic mock inference and a fixture operator; real local tests and persisted receipts. Completed results are identified as recorded attempts. Demo metrics are not production benchmarks.

For “build a greeting function with tests and documentation,” each step has a separate purpose:

What can Accension do?

Tasks can fail, block or need human intervention. Costs and model quality depend on your task and providers. Extra planning, checking and repair can increase spending. Accension does not automatically commit, push or merge changes.

Get started

Manual project and Work Item creation needs no model or paid provider. Use the Windows installer, or install the Python application on Windows, Linux or macOS.

Windows quick start

Requires Windows 10 1809 or later, x64. No separate Python installation or administrator rights are required.

Published Windows downloads are unsigned branch prereleases, not a stable numbered release. They include bundled Python.

1. Download the latest successful main installer and its SHA-256 file.

2. In PowerShell, run the command below and compare its hash with the SHA-256 file. The installer is unsigned; verify its origin and checksum before deciding whether to run it if Windows shows an unknown publisher.

3. Run the installer. Open Accension from the Start menu to open Workbench at http://127.0.0.1:8765/ui/workbench.

4. Choose Accension settings to configure providers, models, routing and repositories. Open Workbench returns to projects.

5. Try the first project tutorial. Use Stop Accension in the Start menu to stop the service.

Get-FileHash .\Accension-latest-windows-x64.exe -Algorithm SHA256

Download and run a newer installer to update; installed applications do not update themselves silently. windows-main follows successful main builds and can change without the package version changing. Check the accompanying source manifest. For uninstall, profile locations, recovery and development builds, read the Windows guide.

Python quick start

Requires Python 3.11+, Git and a source checkout. These commands install that checkout; they do not assume publication on PyPI. Download dependencies once before offline use.

git clone https://github.com/Jigsaw777/accension.git
cd accension
python -m venv .venv

Activate using your shell:

# Windows PowerShell
.\.venv\Scripts\Activate.ps1

# Linux / macOS
source .venv/bin/activate

Then, in either shell:

python -m pip install ".[vault]"
accs init
accs start --background
accs workbench open

init sets up your profile and prompts for repository settings. start runs the local service; workbench open opens the board. These management commands make no model calls and do not edit repository source. You can create projects without a model.

accs status
accs doctor
accs stop

status reports service state, doctor diagnoses configuration, and stop stops the service. Closing the browser leaves it running. If activation fails, use the environment's executable directly; see the complete quick start.

Your first project

This walkthrough makes no inference calls and edits no repository files. It stores project data locally.

1. Open Workbench. Select Personal, open Projects, and save Greeting demo with ticket prefix DEMO. Select it in the Project filter.

2. Open Board → Create work item. Name it Document greeting examples. Set objective: “Show named and blank greeting inputs.”

3. Set acceptance criteria: “Both examples are documented.” Set definition of done: “Examples reviewed.” Save.

4. Open the ticket to inspect its contract. Leave it in BACKLOG while preparing its scope and required checks.

5. Return to Board. The ticket remains after refresh. Follow the Workbench tutorial to add a PRD, dependencies and approved execution.

For a separate disposable demonstration, use a developer installation:

python scripts/demo_workbench.py --feature

This uses mock inference and a scripted fixture operator, runs real local checks, verifies receipts after reopening and cleans up. It calls no paid provider and never uses your working profile.

How the Workbench works

PRD approval accepts requirements. Import creates tasks. Execution approval permits the current scope. Completion review accepts current evidence. Changed contracts or results can require fresh decisions. Failed checks, stale approvals or missing required review prevent DONE.

This synthetic draft makes requirements reviewable before task import.

Review tasks, model policy and spending before starting. Completion remains a separate review.

Full walkthrough · Delivery and recovery · Agent performance · Mobile board

Models, providers, agents and skills

A provider makes models available; a model performs inference. A role describes a responsibility, such as coder or reviewer. An agent is a configured worker. The scheduler coordinates tasks. Skills supply guidance; tools perform authorized actions. Comparison and Android example.

Open Accension settings or run accs ui for model setup at /ui. Connect Ollama or another provider, check capabilities/prices, and choose a routing policy. Accension does not download model weights. Missing evidence or unknown prices can make models ineligible.

Local inference keeps model calls on your machine. Authorized cloud inference can send project context externally even though the database stays local. Repository privacy, provider policy, egress and spending limits all apply. Providers · Routing policy.

Why verification matters

A model reporting success does not automatically mean the work is complete.

For the greeting example, register a repository and trusted tests command, allow greeting.py, and require named/blank inputs to pass. An approved attempt freezes that contract. The Engine applies permitted changes and runs registered checks. A failed test blocks completion; inspect the attempt before retrying.

Real local tests ran with mock model output. Passing tests provide evidence; applicable human review still determines completion.

A receipt records what ran, checks and artifacts. Applicable human review binds to the current evidence. Interrupted workers go through recovery. Allowed paths, protected repository paths and validation constrain execution; they do not guarantee correctness or sandbox arbitrary code.

Delivery behavior · Receipts

Privacy and security

Configuration and history stay local; the service listens on loopback by default. Cloud use is optional and can send permitted context to providers. LAN is opt-in and requires HTTPS, pairing and workspace grants. Remote clients cannot administer provider credentials or repositories.

Credentials use OS storage or explicit environment references. Checks, plugins, skills, credential commands and MCP processes are trusted software. Inspect what you install and authorize: arbitrary extensions are not sandboxed. Allowed paths and validation constrain execution; they do not make model output risk-free.

Privacy · Security and disclosure · LAN setup.

Integrations

Use the browser, CLI, or MCP. Codex and Claude guides explain supported transports. Desktop assistants are not automatically independently controllable Workbench agents. Machine clients cannot grant execution or completion approval.

A human operating the authenticated local CLI can use its explicit approval commands. These are human decisions, separate from ordinary machine API/MCP calls. Feature-batch approval uses the human browser flow. CLI approval reference.

Skills and presets supply reusable instructions; provider plugins extend model access. Validate live provider/client combinations in your intended environment.

What's new in 1.1.0?

1.1.0 adds persistent Workbench planning, delivery and review alongside existing routing. Product 1.1.0, database schema 9, configuration schema 2 and AXIR schema 1 are distinct. Existing CLI aliases and local_ai_router imports remain supported.

Stop active work and back up your profile. Migrations create a verified backup. Older binaries must not open upgraded databases. Restoring a compatible backup loses later database changes and does not undo repository edits. Upgrade procedure · Changelog.

Frequently asked questions

Contributing

Report bugs or request features. Start with development setup/checks and the architecture/source map. PRs require review and CI; passing checks does not automatically merge contributions. Use Security for vulnerability disclosure.

Documentation

The public documentation index groups installation, workflows, routing, security, APIs and contributor references. Technical details stay in canonical guides. Private engineering prompts/task notes stay local.

License

Licensed under Apache-2.0. Copyright 2026 Sourik Ganguly and Accension contributors. Third-party notices.