Documentation
Everything from installing it to signing off a ticket. If you have ten minutes, read “Your first project” and skip the rest until you need it.
npx rekall-accord
Free for seven days. No card, no sign-up.
Accord runs on your own machine. There is nothing to deploy and no account to create.
npx rekall-accord
That downloads Accord, starts it, and opens it in your browser. It prints the address it is serving on and where your projects are stored.
node --version.You can use Accord perfectly well with no coding agent at all. Requirements, baselines, change requests, tickets and QA sign-off all work by hand. The agent is for the building.
In ~/.accord by default. Set ACCORD_HOME to put it somewhere else. It is a SQLite file and some JSON; back it up like anything else that matters.
Click New project. Accord asks one question that changes everything afterwards: is this something new, or code that already exists?
After that, the panel at the top of every project tells you the single next thing worth doing, and the button beside it takes you there. If you only ever follow that panel, you will do the whole workflow in the right order.
In a hurry? Explore sample project on the projects page creates a complete, editable example: an approved baseline, a scope change, tickets, and a feature waiting for QA. It is the fastest way to see the shape of the thing.
This is the feature most people start with. Accord points your coding agent at a repository and asks it to describe what the system does.
You get back an overview of the system and a set of requirements, each with acceptance criteria and the file paths that implement it. Everything arrives as an unconfirmed draft for you to correct. Nothing is agreed until you say so.
Run it again whenever the code has moved on. Accord reconciles with what is already written down rather than starting over: unchanged requirements are left alone, its own earlier drafts are revised in place, and anything you have confirmed or baselined is only ever proposed to you as a change request.
Commit or stash your own work first. A survey only reads, but an implementation run edits files, and you want to be able to tell the agent's changes from yours.
A requirement is a behaviour somebody can agree to. Its acceptance criteria are how you would recognise that behaviour in the running system: the things a tester will actually check.
Three or four criteria is normal. If you find yourself writing one vague criterion to cover a lot of behaviour, you are describing too much at once — split it.
An unconfirmed requirement is a draft. Confirming it is you saying this is agreed. A survey can produce a dozen at once, so there is a Confirm all button; anything without a description or criteria is left behind and named rather than quietly approved.
Guided questions asks the things that are genuinely undecided. On a surveyed project it derives them from what the survey found: areas it did not describe, and places where the documentation and the code disagree. Ask what is unclear goes further and spends an agent run reading the code specifically hunting for what the code cannot settle.
A baseline is a snapshot of exactly what was agreed and when. You can approve one once every requirement is confirmed, has criteria, and every open question is answered. Accord will tell you precisely which item is blocking it.
After a baseline exists, editing a baselined requirement is not allowed. Instead you propose a change, which shows the before and after side by side for somebody to accept or reject. That is what makes “one more thing” visible rather than free.
Accord records your approval and your note. It is not an electronic signature and does not pretend to be one.
Prepare draft PBIs turns confirmed requirements into work. Every requirement produces a pair:
Finishing a build automatically hands its verification to a tester. Sending a build back for more work takes it away again and clears anything recorded against the old code, because that result was about something that no longer exists.
When requirements come from a survey, the work is already built. Those builds are created Done, assigned to nobody, with the survey's citation as their evidence — and only the verification is left outstanding. You will not be handed a backlog telling your team to write the application they already have.
Use the All / To build / To verify filter on the delivery board to look at one half at a time.
Delivery → Execution agent → Choose agent. Accord marks the ones it found on this machine, so you are not offered something that cannot run.
npm i -g @openai/codex, then sign in.npm i -g @anthropic-ai/claude-code, then sign in.ZHIPU_API_KEY on this machine.By default only implementation runs may write, and only inside the project's folder. Surveys and research are read-only. Full access removes the sandbox for every run, lets the agent reach the network and write outside the workspace, and stops the folder binding being enforced. Turn it on deliberately or not at all.
A model tagged -cloud in Ollama is served by Ollama remotely, so your code leaves the machine exactly as it would with any other vendor.
From AI connection:
Accord will refuse an implementation run whose scope is not settled, and will tell you which requirement is unconfirmed or which change is still pending. That refusal is the product working.
The run card shows what the agent is doing in plain words. What changed in the folder above it reads the repository itself, not the agent's account of it, and shows the real diff of every changed file so you review the work rather than a summary of it.
If the agent hits something only you can decide, the run pauses and asks. Answer in the run card and it resumes the same session.
Open the verification ticket. Each acceptance criterion is listed with whatever the agent claimed about it, clearly labelled as the agent's claim. For each one you record Pass or Fail and a note of what you actually did, plus overall test evidence.
Only then can the ticket be marked Done. If any criterion has no result, Accord refuses and says so.
This is the rule the whole product is built around, and it binds the owner too. If you cannot verify a criterion, leave it untested and say why. An honest “not tested” is worth more than a pass nobody checked.
Accord has four roles: owner, developer, QA and client. They decide who may agree scope, who may implement, and who may record test results.
Be aware: this build runs on your machine and has no sign-in, so everyone who opens it is you. Adding people records who is responsible for what, on the specification and in the audit trail. It does not give them access to anything, and no email is sent.
The trial starts the first time you run Accord and lasts seven days. No card, no account, nothing collected.
After that it is $29 a month. Enter your licence key under Enter a licence key at the top of the workspace. A licence covers two machines, so a laptop and a desktop are fine; release a seat before moving to a third.
Reading is always free. Every project, requirement, ticket and test result stays exactly where it is and stays readable, printable and shareable. What stops is making changes. A run already under way always finishes, because abandoning an agent halfway through your codebase would leave a mess.
Accord checks your key with Lemon Squeezy about once a day and then works offline for up to a fortnight at a time.
A failed run shows exactly what the agent's vendor said, and then tells you what to do about it and which other agents are installed on your machine. The common ones:
Your CLI is out of date. Accord shows the update command; for Codex that is npm i -g @openai/codex@latest.
That account has no quota left. Wait for it to reset, buy more, or switch to another agent under Delivery → Execution agent.
Sign in to that CLI on this machine and start the run again.
Check the folder under AI connection is the one you meant, and that the agent you have chosen is one Accord found on this machine.
Email contact@rekall-ai.com with what you were doing and what the run card said. The run's activity log has the detail.