// open source · runs locally

Local tools for
your coding agent.

A desktop app, skills, shared memory, searchable docs, a code graph and a usage dashboard for the CLI you already run: Claude Code, Codex, OpenCode, PI or Oh My Pi.

$ install desktop ▶ watch a prompt travel

5: harnesses
13: tools metered
1: setup for all
0: telemetry

Current Harness Desktop chat: the top bar with Approvals, Memory, Thinking, Cost and Stats, a proposed memory note with Skip, Edit and Save, and three subagents running live

$ harness trace --prompt "Handle empty config input"

The life of one prompt in the family

Follow a single request through the family: the connection in use lights up, and a light runs from the sender to the receiver.

step 01 / 14

You type a prompt

In the Desktop chat of an open project, with the CLI you already use and your own login.

from: you
to: Harness Desktop
sends: your prompt
result: A real window around your harness.

Which tool do I need?

Start with Harness Desktop, which installs the other five for you, or pick one tool: Skills for the workflow, Memory for shared project notes, Knowledge for searchable docs, Code for a code graph, Dashboard for token usage and cost.

Harness Desktop

GPL-3.0-or-later

Your coding-agent CLI in a native desktop window: chat, sessions, review comments, config and extensions.

runs as

Desktop app

platforms

macOS (Apple Silicon and Intel), Linux (AppImage, deb, rpm), Windows (NSIS installer, MSI)

works with

OMP
PI
OpenCode
Claude Code
Codex

data lives

Hooks as JSON files in harness-hooks/hooks in your data folder

Harness Skills

GPL-3.0-only

us-* skills that take a request to reviewed, verified work.

runs as

Plugin or skills loaded by your harness

platforms

macOS, Linux, Windows (PowerShell installer for OpenCode and Codex)

works with

Oh My Pi
Claude Code
OpenCode
Codex
Pi

data lives

Plugin files in your harness

Harness Memory

GPL-3.0-or-later

One shared project memory for your coding agents.

runs as

Local stdio MCP server (one binary)

platforms

macOS (Apple Silicon and Intel), Linux (x86_64), Windows

works with

Any MCP client
Claude Code
Codex
OMP
PI
OpenCode

data lives

memory.sqlite3 in the harness-memory folder of your user data directory

Harness Knowledge

GPL-3.0-or-later

Searchable project documentation with cited passages for your coding agents.

runs as

Local stdio MCP server (one binary)

platforms

macOS (Apple Silicon and Intel), Linux (x86_64), Windows

works with

Any MCP client
Claude Code
Codex
OMP
PI
OpenCode

data lives

knowledge.sqlite3 in the harness-knowledge folder of your user data directory, or a team database you configure

Harness Code

GPL-3.0-or-later

A code graph of your repository: definitions, callers, impact and call chains.

runs as

Local stdio MCP server (one binary)

platforms

macOS (Apple Silicon and Intel), Linux (x86_64), Windows

works with

Any MCP client
Claude Code
Codex
OMP
PI
OpenCode

data lives

code.sqlite3 in the harness-code folder of your user data directory

Harness Dashboard

GPL-3.0-or-later

Token usage and API-equivalent cost across your AI coding tools.

runs as

Desktop app with an optional MCP server

platforms

macOS, Linux (x86_64), Windows

works with

Claude Code
Codex
Gemini CLI
Copilot CLI
Copilot Chat
OpenCode
pi
omp
Zed
Cline
Roo Code
Kilo Code
Cursor

data lives

A local SQLite database on your machine

git diff --stat your-workflow

What changes the day you add the family.

Before:Every session starts cold and rediscovers the repo.

After:Sessions start with the repo’s notes, recalled by any harness. memory
Before:The agent guesses how your billing service works.

After:It searches the docs you imported and cites the passage. knowledge
Before:Grep, open ten files, grep again to follow one call.

After:One code_path answer: the call chain with path:line. code
Before:The agent jumps straight into edits.

After:Research, a plan you approve, parallel workers, fresh reviews. skills
Before:Read the diff in a pager, retype your feedback.

After:Comment on any diff line and send it all as one prompt. desktop
Before:Cost shows up as a bill at the end of the month.

After:Cost per prompt, branch and model, with live plan limits. dashboard
Before:Five CLIs, five config formats, secrets in plain sight.

After:One window for config and extensions, secrets masked. desktop

harness desktop · the heart

Review like a human. Reply like a prompt.

Your coding-agent CLI in a native window, with everything a terminal is bad at.

Comment on the real diffStaged, unstaged and untracked changes. Click a line, then send every comment back as one prompt.
Every session is a branch treeOpen any transcript, resume it, or fork from any point.
Approvals you can readAllow a shell command once, or deny it, before it runs.
One place for five harnessesConfig, MCP servers, plugins and skills of OMP, PI, OpenCode, Claude Code and Codex, with secrets masked.

Read more about Harness Desktop README

5 1 3

⌄ UNSTAGED2

M assets/logo.png +0 −0

M src/lib.rs 1 +15 −3

⌄ STAGED2

R docs/guide.md +1 −1

A src/error.rs +6 −0

⌄ UNTRACKED1

U notes/todo.md +3 −0

$7.88

lib.rs Modified +15 −3 Click a line to comment · Shift+click a second line for a range
1  //! Line-oriented record parser.
2  
- pub fn parse(input: &str) -> Result<Vec<Record>, String> {
+ mod error;
+ 
+ pub use error::Error;
+ 
+ pub fn parse(input: &str) -> Result<Vec<Record>, Error> {
8      let mut records = Vec::new();
9      for line in input.lines() {
10          records.push(Record::from_line(line)?);
⇕ Show 6 unchanged lines
17  }

✓ Review sent to “Handle empty config input”

harness skills

From a request to a verified PR, with a plan you approve.

us-* skills that your harness discovers and follows with its own tools and workers.

Research before editsA read-only scout maps the code, callers and conventions first.
A plan with acceptance criteriaNothing changes until you approve it.
Workers on the model you chooseFrontend, backend and general packages run in parallel.
Fresh reviews, back for repairIndependent correctness and security reviews send findings back to Build.
A release assessment before the PRA read-only assessor grades the change T0–T6 with a confidence and tells you which files to check.

Read more about Harness Skills README

harness memory

Agents that remember your repository.

One local binary, one MCP server, shared by every harness you use.

Search first, save lastFamily hooks make agents look up notes at session start and record what they learned at turn end.
You stay the editorProposed notes appear in the chat with Save, Edit and Skip.
Typed, linked notesDecisions, fixes, problems and locations, linked to files and to each other.
Private by defaultSQLite with full-text search, secrets redacted, no network unless you set up a team database.

Read more about Harness Memory README

harness-memory github.com/useful/harness-desktop MCP · stdio

● new chat · sessionStart

+ Search Harness Memory first

› note_search("")

← 2 notes recalled

› edit .github/workflows/release.yml

timeout-minutes: 10 → 30

● turnEnd · Save what you learned

FixNotarytool needs a 30-minute job timeout

waiting in your Harness Desktop chat

notarytool waits 10 minutes by default and Apple can be slower. release.yml now sets timeout-minutes: 30.

Skip Edit Save

notes · this repository

FixNotarytool needs a 30-minute job timeout

ProblemThe release build fails on macOS when notarytool…

FixTauri window stayed blank after a hot reload

DecisionNotes are read through the Memory CLI

LocationWhere the notes view keeps its state

OutcomeVirtualising the session list cut the first paint

next session, in any harness, starts here

harness knowledge

Agents that read your docs before they guess.

Guides, ADRs, API specs and runbooks you import, searched by every harness through one MCP server.

Cited passagesEach hit comes back with its title, section and snippet, and the agent reads the one it needs in full.
Hybrid searchKeyword and vector ranking merged, with a multilingual model that runs on your computer.
Every format you keep docs inMarkdown, HTML and Confluence exports, OpenAPI, AsciiDoc, reStructuredText, PDFs and meeting transcripts.
You approve what agents addA document an agent proposes waits in a review queue until a person approves it.

Read more about Harness Knowledge README

harness-knowledge github.com/useful/payments-api MCP · stdio

● prompt · refunds retry twice?

+ Check Knowledge first

› knowledge_search(

"")

← 3 pointers · BM25 + vector

› knowledge_get("refunds.md")

Billing › Refunds › Retries

Retries back off 1 min, 5 min, 30 min, then flag the refund.

› knowledge_add → review queue

Reviewrunbook: refund backfill

knowledge_add · waiting for a person

How to re-run refunds that failed in an outage.

Skip Approve

docs · imported

mdhtmlopenapipdfadocrst

runbook: refund backfill ✓ searchable by every agent

docs/billing/refunds.md › Retries Failed refunds retry three times, backing off…

openapi.yaml › POST /refunds Idempotency-Key header required; a retry returns…

adr/0007-idempotency.md › Decision Every write carries a key so a retry is safe…

the next search finds it

harness code

Ask the code graph before you grep.

A tree-sitter index of your repository that answers where code lives, who uses it and how one part reaches another.

A map of the key filescode_map ranks a project’s files with their signatures, within a token budget.
Who uses it, what breakscode_symbol and code_impact list definitions, callers and dependents with path:line.
How A reaches Bcode_path returns the shortest call chain, one hop per line with the call site.
Always fresh, never uploadedEach call re-reads only the files that changed. The index is one local SQLite file.

Read more about Harness Code README

harness-code github.com/useful/harness-cli MCP · stdio

› grep -rn "load_config"

+ load_config is indexed → code_symbol

› code_path("",

"parse_toml")

run_clisrc/main.rs:41

└ let cfg = load_config(&args)?;

load_configsrc/config/mod.rs:88

└ let text = read_source(&path)?;

read_sourcesrc/config/source.rs:23

└ parse_toml(&text)

parse_tomlsrc/config/parse.rs:12

└ pub fn parse_toml(text: &str)

› code_impact(parse_toml)

Impactparse_toml 3 callers · 2 tests

depth 1 · load_config, read_manifest

depth 2 · run_cli

tests · empty_config, roundtrip

code map · ranked

main.rs fn run_cli(args: Args)

mod.rs fn load_config(path: &Path)

source.rs fn read_source(path: &Path)

parse.rs fn parse_toml(text: &str)

args.rs struct Args

lib.rs pub mod config

the index refreshes on every call

harness dashboard

See where every token goes. Down to the prompt.

Thirteen AI coding tools, read from the sessions they already write to disk, stored in one local SQLite file. Never uploaded.

Flow chart of spend from tools to models to projects — Where the money goes · real screen, demo data

Per prompt

What each request consumed, call by call, subagents included.
Per branch

What each piece of work cost, with its commits and pull requests.
Model drift

Whether a model got slower, clumsier or more error-prone than usual.
For agents too

An MCP server lets agents read budgets and plan limits before long work.

Where your data lives: here.

Every tool keeps its data on your machine. Only three connections ever leave it, and you choose all three.

your machine

desktop

harness-hooks/hooks/*.json

No accounts, no stored tokens. Credential files are never read.
memory

harness-memory/ · SQLite FTS5

Per git repository. Secrets redacted before anything is stored.
knowledge

harness-knowledge/knowledge.sqlite3

Imported documents and their index. Agents only propose; a person approves.
code

harness-code/code.sqlite3

The code index of every repository. Nothing is written into a repository.
dashboard

usage.db · read from disk

Reads the session files your tools already write. Never uploaded.
skills

plugin files in your harness

Instructions your harness reads. No workflow executor, no phase database.
your CLI

omp · pi · opencode · claude · codex

The harness you installed, with your own login. It is the only door to your model provider.

Your model provider through your own CLI and your own login
GitHub releases signed update checks, can be turned off
A team memory or knowledge database only if you configure one

Watch them work. About a minute each.

desktop · chat, subagents and review

dashboard · where your tokens go

skills · from request to verified PR

One install. It brings the family.

Install Desktop, then Ecosystem → Run recommended setup. It installs Dashboard, Memory, Knowledge, Code and Skills, registers their MCP servers in every harness it finds and adds the family hooks, asking before each step.

# macOS · Linux

$ curl -fsSL https://raw.githubusercontent.com/MSpiechowicz/harness-useful-desktop/main/install.sh | sh

# Windows

> powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/MSpiechowicz/harness-useful-desktop/main/install.ps1 | iex"

Questions about the family

What is Harness Useful?

Harness Useful is a family of six open-source, local tools for coding-agent harnesses such as Oh My Pi, PI, OpenCode, Claude Code and Codex. Harness Desktop is the window, Harness Skills the workflow, Harness Memory the shared project memory, Harness Knowledge the searchable project documentation, Harness Code the code graph and Harness Dashboard the usage and cost meter. Each works on its own.

Which Harness Useful tool should I start with?

Start with Harness Desktop. Install it, then open Ecosystem and choose Run recommended setup. After you confirm each part, it installs Harness Dashboard, Harness Memory, Harness Knowledge, Harness Code and Harness Skills, registers their MCP servers in each detected harness and adds the family hooks. You can also install any tool by itself.

Do the Harness Useful tools work separately?

Yes. Each tool has its own repository, installer and README, and works without the others. Harness Memory, Harness Knowledge and Harness Code are standalone MCP servers, Harness Dashboard is a standalone app, and Harness Skills installs straight into your harness. They work best together, and Desktop connects them through Recommended setup.

Which coding agents do the tools support?

Harness Desktop supports Oh My Pi and PI, with partial OpenCode support and chat previews for OpenCode, Claude Code and Codex. Harness Skills supports Oh My Pi, Claude Code, OpenCode, Codex and Pi. Harness Memory, Harness Knowledge and Harness Code work with any MCP client. Harness Dashboard reads 13 tools, including Gemini CLI, GitHub Copilot, Zed, Cline, Roo Code, Kilo Code and Cursor.

All questions and answers