For teams already shipping real work with coding agents
Your agents do the work. You decide what moves forward.
Attesta is a local desktop app for teams that build software with coding agents. Today the same chat that wrote a change is asked whether the change is fine — and when it ships, all you hold is a transcript. Attesta runs your process around the CLIs you already use: a person decides what moves forward, and the record writes itself while the work runs.
Illustrative walkthrough with sample data — one Line of many. Your project composes its own stations and their order, with a human key wherever you put one.
01 / The team
Coding tools are not made for teams.
One person, one chat, one machine — that is the shape of every agent tool today. Run a team on that shape, and the same failures arrive on schedule.
Today · without a process
The team’s knowledge never reaches the session.Rules, standards, and decisions live in heads, wikis, and pastes — they arrive only if someone remembers to bring them.
Context pollution degrades the work.Sessions accumulate stale plans, dead ends, and yesterday’s rules — and a polluted context degrades an agent’s output.
Everyone ships on their own terms.Their own prompts, their own checks, their own definition of done. The team’s process is whatever each person happened to do.
The process gets locked to one vendor.The only way to get consistency today is to standardize on a single tool — then the prompts, rules, and history live in its format, on its terms, and switching means starting over.
With Attesta
Attesta gives your team a process.
Declared by the project, carried by the repository — it reaches every seat, not just whoever remembered. People change seats, agents change names; the process holds.
A change — a Job — rides a Line your project composes: the stations, their order, and a human key wherever you put one. Attesta advances governed work only when each station’s declared conditions are true. The map below is a different Line from the window above; a project keeps as many as the work needs.
Briefpersonintent · constraints
Planagentbrief · context
Validateindependentplan · rules
Buildagentplan · checks
Reviewindependentdiff · evidence
Approvalpersonexact evidence
Rules freeze when work starts.A later edit governs a later Job; it does not rewrite the basis of work already underway.
Authorship and review can be separated.Where the Line requires independence, an agent already on the record is refused.
Human authority is explicit.A named decision covers the evidence actually shown — not a generic green state.
03 / Context
You don’t need a magic prompt or a new skill. Your agents need the right context, served.
A capable model that fails on real work is usually starved, not stupid. Durable project knowledge lives outside disposable chats — what to build, what it must look like, what the team already decided and why — and each station is served a bounded set of it. The serve is recorded.
Rulebookhow agents behave
One source of working rules, kept consistent for every agent CLI you seat.
Design systemtables
What the result must look like — the exact canon this station needs.
Memory2 facts
Sourced facts that outlive the agent, the chat, and the employee. They inform — they never approve.
Repository mapbounded
When the shelf has no answer, Attesta reads only the repository slices this station needs — never a whole-repo dump.
Plan station — served
The agent opens with the declared task and its exact, inspectable context.
Attesta does not claim that a badge makes software safe. It records the concrete basis a team can inspect: which process governed the change, who acted, which checks ran, what evidence existed, and what a person approved. That book is the Ledger — append-only, a change’s place is the sum of its rows. It lives with the project, on your machine. You hold the book.
021EVIDENCE3 files · byte fingerprints · viewed set 6af1…921c
022APPROVEDm.chen · exact evidence · named human act
Approval basis
row 022 · inspected before record
Frozen LineStation order and instructions fingerprint match.
Independent reviewThe reviewer does not appear in the author set.
Checks witnessedResults came from Attesta running the checks — not from the builder’s claim.
Human decisionA named person approved the evidence actually rendered.
05 / The Seal
Enforced, not promised.
Everything above is a mechanism, not a convention — and the proof is the Seal: a verdict computed by replaying the book. No judgment, no trust. When every gate ran, every approval is named, and every evidence byte matches, the Seal stamps. When it doesn’t, red names the exact row.
Replay — change 042 · csv_export41 rows · sample
✓Line frozen at open — fingerprint 7c2e·88a1 · approved basis rev 12
✓stations ridden in order — Plan → Validate → Mockups → Build → Review → Approval · 6 of 6 reported
✓evidence re-hashed — 9 files, byte for byte, all match their rows
✓human keys where the Line demands them — every approval named
✓Line frozen at open — fingerprint 91d4·c3aa · approved basis rev 12
✓evidence re-hashed — 6 files, all match their rows
✗Review reported with no independent session on the record — row 29 names it
no seal — the route was not ridden
A verdict, not a badge.The Seal is computed from the book — it can only be replayed, never granted. No one can manufacture one by hand.
A skipped gate stays visible.There is no override that writes “sealed.” The only path to the stamp runs through the stations.
Your repo can demand it.Make the Seal a required check, and only work that rode an approved Line merges.
06 / The floor
The project is the durable unit — not the chat.
The repository carries the Lines, the rules, and the book — files in your repo, runtime state on your machine. People and agents can change; the project remains.
billing-exports-prdProduct record · kept versions · sample
v1v2v3v4
v4 — exports above 50,000 rows queue. Ruled by a person at the Validate station; the losing words stay on the record. A kept version is immutable.
sourced · informs every later serve · never approves
Your subscriptions, your machine. Attesta runs on the agent plans you already pay for — it does not broker model access or resell tokens.
Solo needs no Attesta account. The floor is the same with one seat open or ten.
A decision made once is not made again. Rulings become kept versions and Memory facts that every later station is served.
07 / Early access
Your agents build. You decide. The record holds.
Attesta is opening to a small number of teams already shipping real work with coding agents. There is nothing to install today — leave a name and a work email, and a person writes back. It runs on the agent plans you already pay for; the book stays on your machine.