rsc-harness · memory for coding agents

Talk to your agent. Watch it remember.

claude code · ~/my-appsession 1
›
02-DOCS/0 pages0 links
drag to orbit
Help us with a star on GitHub

Free and open source. A star is how the next developer finds it.

get started02

You build the product.rsc builds its harness.

No config files, no setup guide. You send one message to the agent you already use, and it sets itself up.

1 Paste this into your agent Claude Code · Codex · Cursor · Gemini CLI · and 14 more

Read https://ericrisco.github.io/rsc-harness/ and set up rsc for this project. Ask me the onboarding questions, show me the exact plan, and wait for my acceptance before writing.

or in a terminal npx @ericrisco/rsc@latest onboard
then, in your agent
  1. 2

    It asks a few plain questions

    How technical to talk, and what you are building. It reads only the project root.

  2. 3

    It shows you the exact plan

    What it installs, what waits for evidence, what it leaves out. Each with its reason.

  3. 4

    You say yes. Then it writes.

    Nothing touches your project before that. A handful of skills, not all 273.

18 assistants273 skills, a handful installedMIT, no API key
~/my-project · rsc onboard
Example plans. Pick a kind of project and watch the plan change.
what you get03

One message installsfour things.

01memory02-DOCS/

It remembers your project.

A wiki of your decisions, stack and progress. The agent reads it before it starts and adds to it as it works.

02hands01-TOOLS/

It works with your tools.

Your CRM, your ERP, your APIs. Each connection is built once, tested, and kept for every next session.

03tradeskills

It knows your stack.

The know-how for what you build, picked for this project. A handful out of 273, so the rest never costs context.

04reflexeshooks

It checks before it says done.

Rules that run by themselves: proof for every task, the right amount of process, your language in every reply.

why documentation comes first04

Searching costs tokens.Finding does not.

An agent without documentation opens files, greps, reads the same module twice, and still guesses. With a well-structured 02-DOCS/ it knows where to look.

Without docssearching
0tokens read
With 02-DOCS/finding
0tokens read

Illustration of the mechanism, not a benchmark.

No duplicate codeIt reuses the function you already have instead of writing a second one.
One structureIt follows the patterns your project already uses, every time.
Less that breaksIt works from what is true about your code, not from a plausible guess.
the hands, up close05

Name a tool.It gets connected.

CRMs, ERPs, the program on the office PC. connect-tool reads the vendor's current docs, takes the safest route, builds 01-TOOLS/ and proves it works. Pick one.

1
read the docs
Pick a tool on the left.
2
safest route
 
3
build
 
.env.exampleCREDENTIALS.mdREADME.mdtest_connection.sh
4
prove it
 

Example run. The skill checks each vendor's current docs every time, and says so when it finds nothing reliable. Any tool works, not only these.

a day with rsc06

Tomorrow,it already knows.

Five steps, from install to the next morning.

install

A few plain questions

How technical to talk, what the project is for, its goal, which assistants. It reads only the project root.

? How should I talk to you?
› technical
the plan

Nothing is written until you accept

Every piece comes with its reason. You accept that exact plan, or nothing happens.

selectedorient, ftd, connect-tool
deferredasdd, until there is evidence
excludedvideo, payments
work

It picks the right amount of process

An answer, a quick change, or the full chain. It says which in one line, in your language.

lane: FTD · one feature document
write down

It documents as it goes

Decisions, connections and progress land in 02-DOCS, not in the chat.

02-DOCS/wiki/ftd/csv-export.md + evidence
next morning

It continues where it stopped

You, a teammate on another machine, or another assistant open the project, and the agent already knows.

next: verify the export with real filters
three lanes07

The right amount of process.Chosen for you.

read-onlyAnswerExplain, compare, audit. Writes nothing.
defaultFTDOne feature document, proof for every task.
big or complexASDDThe agent enters it on its own. You approve the spec.
safe for teams08

What one person teaches,everyone's agent knows.

  • Knowledge reaches everyone01-TOOLS and 02-DOCS travel through one exchange branch, and arrive in main inside normal pull requests.
  • One workspace per agentA second session in the same folder works in its own worktree, removed once the branch lands.
  • No surprise commits on mainYou choose at install: straight on main, or branches and pull requests.
  • Any assistant, one installPick several. The skills are written once and shared.
the catalog09

273 skills.You install the few that fit.

8 of 273the rest stay out of your context
Hover a square to see the skill. Each one fires on its own when a task needs it.
why you can trust it10

Evidence over claims.

P10

Done means observed

A task is checked off when its proof was seen: the test output, the command's result. Belief does not count.

P2

Gates that can fail

Every rule rsc declares binding ships with the check that enforces it, and a test that watches that check fail and pass.

P7

Friction where it matters

Silence for the trivial, ceremony only where failure is expensive. A harness that nags gets switched off.

Known limits, stated openly: the hard guards run in Claude Code; on other assistants the rules travel as instructions. Session memory is local by design and never touches the network.

for AI agents reading this page

Build the user’s harnessthrough the binding wizard.

Do not choose components or onboarding answers for the user.

agent-instructions
  1. Work in the user’s project root, never in a clone of rsc-harness.
  2. Run npx @ericrisco/rsc@latest onboard and relay every required question.
  3. Non-interactive fields include --technical-level --project-kind --goal --software-scope --target.
  4. Show the selected, deferred and excluded plan. Never invent answers.
  5. Rerun with --accept-plan <printed-id> only after exact acceptance.
  6. Only RSC_ONBOARDING_READY means the harness is complete. Tell the user to reload the session.
  7. RSC_ONBOARDING_INCOMPLETE means it applied but the harness floor is missing: run the action it names, and never tell the user rsc is ready until it prints READY.
questions

Questions, answered.

What is rsc-harness?

rsc-harness is a free, open-source (MIT) harness builder for AI coding agents. One message to Claude Code, Codex, Cursor or another of its 18 supported assistants installs a project memory (02-DOCS/), tested tool connections (01-TOOLS/), a handful of skills chosen from 273, and hooks that check the work before the agent says it is done.

What is an agent harness?

Everything around the model that makes it useful in a real project: the instructions it reads, the memory it keeps, the tools it can call, the skills it loads and the checks that run on their own. The model is the engine. The harness is the rest of the car.

How do I install it?

Paste the setup message into your agent, or run npx @ericrisco/rsc@latest onboard in the project root. It asks a few questions, shows the exact plan (selected, deferred and excluded pieces, each with its reason) and writes nothing before you accept it. Agents can run it non-interactively with --technical-level and, after your acceptance, --accept-plan.

Which AI assistants does it support?

Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, DeepSeek Harness, Windsurf, Cline, Roo, Zed, Continue, Amp, Antigravity, Jules, Junie, Kiro and Aider. You can pick several: skills are written once and shared.

Why not install every skill?

That is how a Frankenstein harness happens: duplicated skills, premature hooks and unnecessary MCP servers that fill the context and contradict each other. rsc installs only what the project root and your answers justify, loads skills progressively, on demand, and adds more only when new evidence appears.

Does good documentation really lower token cost?

Yes, as a mechanism. An agent with a well-structured 02-DOCS/ knows where to look instead of searching file by file, so it reads less, reuses the code you already have instead of duplicating it, and guesses less. We show the mechanism, not a benchmark number.

Is it free, and does it send my code anywhere?

It is MIT-licensed and needs no API key and no account. Session memory is local by design and never touches the network. It runs inside the assistant you already use.

Does it work for teams?

Yes. Knowledge in 01-TOOLS/ and 02-DOCS/ travels through one exchange branch and reaches main inside normal pull requests, a second agent in the same folder gets its own worktree, and you choose at install whether work goes straight to main or through branches and pull requests.

start here

Stop explaining.Start building.

$npx @ericrisco/rsc@latest onboard