# your team’s judgment, wired into every agent
Your AI writes code like it
just joined the team.
Every single session.
Stop re-explaining your team's decisions and how you work. Context Engine gives your coding tools the decisions behind your code. Memory carries your working preferences into the next session. Worker takes on engineering tasks with that context. Start with the product that solves your problem; each works on its own.
Context Engine and Memory are live. Sign up while seats are available; join the waitlist when full. 50 free credits, no credit card. Each plane installs on its own and is worth running alone.
# 01 · what it costs you
Every agent you run is a new hire with no onboarding.
It has no memory between sessions and no route to the reasoning behind your codebase, so it does not stop and ask. It makes a locally sensible change that contradicts something your team settled months ago, and the cost lands on the people reviewing it.
- 0132.7%of AI pull requests are accepted, against 84.4% of the ones a person wrote
- 025.3×longer to review an AI pull request than a human one, and 4.6× longer waiting
- 035k to 20ktokens burned every session re-establishing what the codebase already knows
- 04~1 dayper developer per month, spent re-explaining context you explained last week
Acceptance and review figures from the LinearB 2026 engineering benchmarks. Token and time figures are the range reported across published 2026 developer surveys.
# 02 · the record · decispher context engine
Stop hand-writing the file that tells your agent how you work.
Every team is now maintaining some version of it: CLAUDE.md, AGENTS.md, .cursorrules, a wiki nobody opens. The file is not the problem. Hand-written rules go stale the moment somebody changes their mind in Slack, and nobody goes back to edit them.
Decispher reads the decision where it was actually made, in the thread, the review or the ticket. The same call stated three times is one unit, not three, and agreement across independent sources is what raises its confidence.
“we should never update ledger rows, just insert a reversal”
“reverted: this UPDATE breaks the audit trail”
“ledger is append-only, corrections are new rows”
credit_ledger is append-only. Never update rows in place.
one source is an opinion. three independent sources, across three tools, is a rule. 0 conflicts · 0 supersessions.
then written into every file your tools already read
Opened as a pull request against your repository, regenerated when the record changes, and enforced in CI by Decision Guardian, which is open source. Agents that speak the Model Context Protocol skip the files and query the record live through 25 tools.
# 03 · memory · decispher memory
You already told it. Twice. In two different tools.
Use pnpm, not npm. Staging is on 5433. Do not touch the generated client. You said it in Cursor on Monday, again in Claude Code on Tuesday, and you will say it again on Thursday.
Say it once to Decispher and it arrives before your next prompt, on every tool and every machine, scoped to you, your project or your whole company.
decispher · injected here, by us
3 served · receipt written · yours to delete
wire up the new billing table
- 01Three scopes, one ruleYours, your project’s, your company’s. Company policy always applies; under it the most specific wins.
- 02A receipt per serveWhich memory, to whom, in which session, at what token cost. Exportable, and no receipt means no serve.
- 03Yours to deleteImmediate and everywhere. Only a bodyless receipt survives, so the audit trail never gains a hole.
- 04Never inside your promptInjected at seams Decispher owns, before your message. Your words reach the model exactly as you typed them.
# 04 · the worker · decispher worker
Two in three AI pull requests get sent back. Rarely for syntax.
They get sent back because the change is locally sensible and globally wrong: a deprecated helper, a pattern the team retired, a table somebody decided was append-only. Every AI engineer you can hire today reads your repository and guesses your judgment from it.
This one is briefed from what your team decided, works in a sealed sandbox on its own branch, and stops when the task and the record disagree.
# 05 · take what you need
Three verticals. One bill.
No package deal.
Each plane installs on its own and is worth running alone. They share one tenancy model, one credit balance and one database, and are deliberately unlike each other everywhere else, so you can start with whichever one hurts most today and add another when it earns its place.
There are no per-plane plans and no seats. Every action has a published cost, spend is capped per project and per mode, and a Worker run reserves its budget before it starts. You pay for what you actually use.
# 06 · the receipt
A context product you can audit instead of trust.
Everything in this category is easy to claim and hard to check. So Decispher receipts itself: every serve is logged with what it cost, every run states which decisions it was grounded in, and the savings are measured per call against what that call would have cost cold, never estimated once in a slide.
The same discipline runs underneath: no receipt, no serve. A memory that was injected is logged, and one you delete leaves a bodyless receipt behind so the audit trail never gains a hole.
how we measure ↗# 07 · start
One command. Nothing about how your team works changes.
npx decispher initIt wires this repository for the agents you already run, writes the shells they read, and starts recording sessions on the branch you are on. Connecting Slack, GitHub or GitLab and Jira is a few minutes of OAuth in the dashboard. Nobody has to change where they talk, review or plan.
works with, on day one
Raw Slack and Teams messages are never stored beyond the analysis window. Every context unit is encrypted at rest under a key derived per company. Capture shows a recording light and you can pause it at any time.
# founding cohort · early access open
Give your agents the last three years of your team’s thinking.
We onboard founding teams by hand, because the first week decides whether a context product is worth keeping. You get the setup call, direct access while you evaluate, and pricing locked for the life of the account.
No credit card · drop-in MCP server · pause capture any time
Context Engine and Memory are live. Worker is early access, enabled per company while we onboard founding teams.
Before you sign up.
- What is Decispher?
- Decispher wires your team’s judgment into every AI agent you run, and it is three products that each install on their own. The Context Engine is the system of record: it captures the decisions, conventions and constraints your team makes in Slack, GitHub, GitLab and Jira, and writes them into the agent files your tools already read. Decispher Memory remembers the durable working preferences of each person, project and company, so you stop repeating yourself to every tool. Decispher Worker is an autonomous engineer briefed from both, which opens draft pull requests with receipts and never merges.
- Do I have to adopt all three products?
- No. The Context Engine, Memory and Worker are separately adoptable, and most teams start with one. They share one record underneath, so adding a second is a switch rather than a migration, and the value compounds because the record briefs the work and the corrections feed back into memory.
- What does my team have to change?
- Nothing about how engineers work. Decisions are extracted from the Slack messages, pull request descriptions and Jira comments your team already writes. There is no document to maintain and no capture step to remember. The setup is connecting your sources once, and one command in each repository.
- How long does setup take?
- Connecting Slack and GitHub takes a few minutes in the dashboard. Wiring a repository is a single command, npx decispher init, which installs hooks for Claude Code, Cursor, Codex and Grok Build. Pointing an agent at the MCP server is one entry in its config file. Your first session is recorded within a couple of minutes.
- Which AI coding tools does Decispher work with?
- Claude Code, Cursor, Codex, GitHub Copilot, Windsurf, Cline, Grok Build and anything else that speaks the Model Context Protocol. Decispher also generates and keeps in sync nine agent instruction formats, including CLAUDE.md, AGENTS.md and .cursorrules, so tools that read a file rather than call a server are covered too.
- Does Decispher store my source code?
- No. The record stores decisions, reasoning and pointers to files, not your repository, and the platform never clones your codebase. The single exception is a Decispher Worker run, which works on a copy inside an isolated sandbox that is destroyed when the run ends.
- How is this different from a CLAUDE.md file or a wiki?
- A rules file is loaded whole on every call, drifts as the codebase moves, and has no way to mark an entry as superseded. A wiki requires somebody to write and maintain it, which is the step that always stops. Decispher captures from work that is already happening, tracks whether a decision is still current, serves only what a task needs, and flags a pull request that contradicts something on record.
- What does it cost?
- Credits, not seats. New teams start with 50 free credits, and packages begin at 9 dollars for 100 credits. You are charged when a decision is actually extracted or when a tool call runs a model. Detection, memory injection and memory lookup are free, and Worker runs can be billed to your own model provider key instead.
- Is there anything open source?
- Yes. Decision Guardian, the enforcement layer that blocks a pull request contradicting a decision on record, is open source and free to run in CI. It has been through more than 1,300 production runs across 24,000 files with entirely organic adoption, and it has its own section in the canonical architectural-decision-records repository.
- Which products are available today?
- Context Engine and Memory are live and independently usable. Sign up while seats are available; when the cohort is full, join the waitlist. Worker is in early access and must be enabled for your workspace.