Accension

Route coding tasks across local and cloud models, with explicit budgets and independent checks.

Working Python V1: a localhost model gateway plus a bounded repository planner, task DAG executor, validation/repair loop, SQLite budgets and cache, and stdio MCP server. The Python module and installed MCP integration retain the name local_ai_router / local-ai-router.

Try it without cloud credentials: the demo creates a greeting module, documentation and tests in examples/demo-repo, then executes two real Python tests using deterministic mock inference. Mock mode implements this fixture only; it never silently replaces an unavailable real model.

Quick start

Requires Python 3.12+ and a source checkout. Windows PowerShell:

git clone https://github.com/Jigsaw777/accension.git
cd accension
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[test]"
.\.venv\Scripts\python.exe -m local_ai_router.cli demo

On Linux/WSL, run sh setup.sh, then .venv/bin/python -m local_ai_router.cli demo. Use an editable source checkout; standalone wheel distribution is not supported yet.

Real providers start disabled. Configure endpoints, prices and actual deployment IDs in ignored config/local.yaml. Entire YAML sections are replaced, not deeply merged. Azure inventory, local inventories, Laya and personal skills are optional. The router automatically accommodates new deployment IDs on supported protocols; unknown prices and unverified capabilities do not authorize spending.

Bring your own local models: local configuration example and provider/classifier setup. Qwen and Laya are not required.

To start the gateway, run python -m local_ai_router.cli serve inside the virtual environment. It binds to http://127.0.0.1:8765. For generated Windows helpers, use setup.ps1 below. RTK prefixes in the remaining examples are optional: omit rtk proxy if RTK is not installed.

Related MCP server: coder-worker-mcp

Commands

rtk proxy .\scripts\router.cmd status
rtk proxy .\scripts\router.cmd doctor
rtk proxy .\scripts\router.cmd route "Rename this variable."
rtk proxy .\scripts\router.cmd discover
rtk proxy .\scripts\router.cmd config validate

For a real repository, register its exact root and independent test/build commands first. Put the following section in ignored config/local.yaml:

repositories:
  - path: C:/path/to/project
    allow_cloud: false
    validation:
      tests: ['{python}', '-m', 'pytest', '-q']

Then use an eligible configured model:

rtk proxy .\scripts\router.cmd run "Add feature X" --repo C:\path\to\project
rtk proxy .\scripts\router.cmd plan "Add feature X" --repo C:\path\to\project
rtk proxy .\scripts\router.cmd execute PLAN_ID --repo C:\path\to\project

plan persists a typed JSON DAG and readable plan without editing source. execute rejects a stale or already executed plan. run performs both. Existing unrelated edits stay in place; proposed edits require original content hashes. Set allow_cloud: true only for repositories allowed to send source to configured cloud providers. Restart running services after configuration edits.

Installation and development

setup.ps1 creates .venv, installs dependencies, generates local helpers, previews client changes and starts the gateway. Pass -InstallClients to apply backed-up Codex/Claude integrations, user PATH and per-user autostart. setup.sh installs the CLI on Linux/WSL; Linux client registration and autostart are manual.

rtk proxy powershell -NoProfile -File .\setup.ps1
rtk proxy .\.venv\Scripts\python.exe -m pytest -q

The test suite uses mocks and local subprocesses, not paid models. GitHub Actions runs it on Windows and Linux with Python 3.12. requirements-lock.txt records a Windows development environment; pyproject.toml defines supported ranges. There is no Redis, broker, container requirement, frontend build, or agent framework. SQLite is sufficient for one workstation.

Where things live

Read ARCHITECTURE.md, ROUTING_POLICY.md, PROVIDERS.md, CACHE.md, MCP.md, CODEX_SETUP.md, CLAUDE_SETUP.md, SECURITY.md, and TROUBLESHOOTING.md.

V1 boundaries

Native protocol forwarding preserves tool calls and SSE without translating between Responses, Chat Completions, and Anthropic Messages. Workers propose bounded complete-file replacements; the host runs registered checks. Repository tools/tests execute as the current OS user, so use trusted repositories. Semantic cache, Redis, worktree merging, general shell agents, full Kotlin/Java semantic benchmarks, crash-resume of partially applied runs, and billed-cost reconciliation are not implemented. Gateway compatibility is covered by mock protocol tests; a live Codex/Claude inference conversation has not been verified.

Contribute and license

This is an early project. Bug reports, documentation improvements, tests and focused pull requests are welcome. See CONTRIBUTING.md and SECURITY.md.

Licensed under Apache-2.0. Copyright 2026 Sourik Ganguly and Accension contributors. Dependencies retain their own licenses; see THIRD_PARTY_NOTICES.md. No API keys, personal configuration, private reports, model weights or third-party skill content are included.