OLLAMA WORKSPACE AGENT / LOCAL BY DESIGN

A coding agent that works where your code lives.

OwA runs in your terminal with Ollama. It reads the workspace, finds relevant code, makes reviewed changes, and checks the result with real tools.

Start with OwA See the workflow

The new MCP, sandbox, and memory work is on main. Check PyPI for the latest packaged version.

~/projects/calculator › owa LOCAL SESSION

OwA · Ollama Workspace Agent

you › Fix add() and run the tests.

01read_filecalculator.pydone

02patch_filecalculator.pyapproved

03run_commandpython -m unittest -qexit 0

Change verified

add() now returns the sum. The requested tests passed.

Example workflow · actions require your approval

model qwen2.5-coder:7bworkspace local

YOUR OLLAMA MODELYOUR WORKSPACEYOUR APPROVALVERIFIED OUTPUT

01 / THE WORKFLOW

From question to checked change.

OwA keeps the work visible. The model gets focused context and tool results, while you remain in control of actions that change files or run commands.

Understand the project

Read files directly, inspect Git, or build a local index for code search. Evidence is trimmed to fit smaller model contexts.

/index

Plan and edit

Ask for a focused patch. OwA checks workspace paths and asks before it writes. Optional manager and tester roles can review the work.

patch_file

Run and verify

Approve a command, see its output and exit code, then let OwA report whether the requested check actually passed.

run_command

Keep useful context

Local conversation history, the code index, and completed file-change episodes help the next task start with context.

.owa/

02 / THE IDEA

Built brick by brick.

OwA starts with your local Ollama model. Each file read, reviewed edit, and test adds a useful piece. The work stays visible as the project takes shape.

01 Read the workspace 02 Review the change 03 Check the result

Ready to install? Find OwA on PyPI

03 / WHAT IS INSIDE

A small, inspectable toolchain.

The agent uses concrete tools to inspect, edit, test, and explain work in your repository. Optional integrations stay explicit.

WORKSPACE

Know the code before changing it.

Hybrid code search, direct file reads, project orientation, and Git status, diff, and log give the model grounded evidence.

search_coderead_filegit_diff

MODEL

Built around local models.

Native Ollama tool calls, bounded context, and explicit verification help 7–9B models handle focused tasks.

INTEGRATIONS · MAIN

MCP when you ask for it.

Connect optional stdio tools and approve discovery and calls. MCP servers are chosen and configured by you.

MEMORY · MAIN

Remember completed work.

SQLite episodes record completed file changes locally; they are treated as historical context, not instructions.

ROLES · MAIN

Bring in a second pass.

Optional manager planning and tester review surround the main coder loop without hiding the tool activity.

04 / STAY IN CONTROL

Permission is part of the workflow.

OwA checks paths before writes, asks before file changes and shell commands, and can run shell commands in a constrained Docker container. The Docker image must be present locally.

01 Approval gates for writes, commands, and MCP tools
02 Docker command sandbox by default on main
03 Tool results and command exit codes stay visible

MCP servers run with your host privileges. Configure only servers you trust. Host command execution is an explicit opt-in.

OWA / ACTION REVIEW02 OF 03

REQUESTED ACTIONrun_commandpython -m unittest -q

EXECUTIONDocker sandbox

No network · limited resources · workspace mount

AllowDeny

Illustrative prompt. OwA asks in the running terminal.

05 / HONEST EXPECTATIONS

Useful with small models. Still a tool you supervise.

Earlier disposable-project evaluations passed 36 of 36 runs for Qwen 7B and 36 of 36 for Ornith 9B. Those results cover focused fixtures, not every real repository. The latest expanded live run is pending while the configured Ollama server is unresponsive.

Read evaluation reports

06 / GET STARTED

Open a terminal. Bring your own model.

Install the published package for a stable starting point. The newest development capabilities are documented in the GitHub source until the next package release.

INSTALL & RUNbash

pipx install ollama-workspace-agent
cd your-project
owa

# In OwA, type /setup to connect Ollama

01 Install Ollama and a chat model.

02 Run owa inside your project and use /setup to configure the connection.

03 Ask a concrete question or request an edit, then review each proposed action.

Read the full setup guide Check the published package version ↗