A review that survives a model swap
A workstream's scope, evidence, and approval history stay attached to the GitHub issue that authorized it — regardless of which engine ran which step.
An operating layer for durable AI work
AI execution changes.
Your work shouldn’t have to start over.
Agents get swapped. Models get upgraded. Providers change terms. Cortxt keeps the mandate, state, decisions, and evidence attached to the work itself — so none of that turns into a reset.
One workstream. Two runs. Same mandate.
Run 1 executed on Hermes and stopped on a fail-closed policy block. Run 2, after a real fix, executed on claude-direct — same GitHub-issue workstream (#402), same mandate, same evidence trail.
Run 1 · Hermes · blocked→Run 2 · claude-direct · done
Continuity proof write-up — coming soon Workstream authorityLive contract
Desired outcomeShip reviewed policy evidenceWS-042 · Human owned
MandateSignedscope + budget fixed
DecisionRequiredreserved for human
Evidence12 recordsattributable
PolicyFail closedprovider constrained
Run 01 · HermesRun 02 · Codexworkstream persisted
The problem
Every agent session ends. The work it did shouldn’t disappear with it.
Switch models and you lose the thread. Swap an agent runtime and you lose the history. Most tools tie your scope, your approvals, and your evidence to whichever session happened to produce them. When that session ends — intentionally or not — you’re rebuilding context from scratch.
- Context lives in a chat window, not a record.
- Approval lives in someone’s memory, not a mandate.
- Evidence lives in a transcript, not something you can hand to a reviewer.
Durable authority. Replaceable execution.
What must survive shouldn’t depend on which engine you’re on.
Cortxt separates what must survive from what shouldn’t matter which one you’re on. The mandate, the workstream state, the decisions made, and the evidence collected are Cortxt-owned and durable. The engine that executes a given run — Hermes, Codex, or another adapter behind the same contract — is a replaceable resource underneath that record, not the record itself.
Replaceablehermescodex+ more adapters
->
Cortxt contractmandatestatedecisionsevidence
Mandates, made concrete
Give AI work a mandate — not a blank check.
A mandate is scope, limits, and explicit authority, issued before a run starts. It’s not a system prompt and it’s not a permissions checkbox — per ADR-032, every mutating (Tier-1+) MCP tool call requires a signed, nonce-bound mandate envelope, verified before that call executes, and fails closed when the call steps outside it.
In the proof run above, a policy gate rejected an out-of-scope action automatically, before any human had to notice or intervene. That’s what “mandate” means in practice.
Cortxt OS
Workstreams, not agent fleets.
Cortxt OS is the human-facing surface for operating durable work — not a terminal, not a raw agent cockpit. It’s organized around workstreams you have a mandate over, the decisions waiting on your attention, and the evidence a run produced.
Early preview · In development
Work3 need attention
Example workstream A
Example workstream B
Example workstream C
DecisionsWaiting on you
Approve mandate renewal
Review policy exception
Illustrative presentation of the merged first slice with placeholder workstreams — not real customer data or a completed case study.
Individual widget views can be generated from a prompt today (cortxt widget generate). The first Studio describe/generate/confirm flow is now merged; broader adaptive composition remains in development.
Privacy by restraint
A product about durable authority shouldn’t quietly profile its visitors.
The public Cortxt site has no advertising, tracking pixels, visitor analytics, account system, or marketing-cookie layer. The theme switch stores one preference in your browser’s local storage. Hosting infrastructure still receives ordinary request metadata needed to serve and protect the site.
Read the privacy & trust statement ->01No tracking cookiesNo consent theatre for analytics we do not run.02No account requiredRead the site and clone the project without signing up.03Explicit external boundariesGitHub and hosting links have their own privacy terms.
Illustrative
Built for work that outlives any one session.
Long-running work that spans engines
Start on one adapter, continue on another after a provider change, without losing the mandate or the decision trail.
An audit trail by default
Every action is evidence-backed and mandate-scoped from the first step, not bolted on after the fact for a postmortem.
Where things stand
What’s real today, what’s direction, and how to work with us.
| Layer | Status |
|---|---|
| Mandate issuance & verification, fail-closed policy gate | Available now |
| GitHub Issues as workflow / evidence authority | Available now |
| cortxt CLI, incl. cortxt mcp serve | Available now |
| Engine adapters (Hermes, Codex, and more behind the same contract) | Available now |
| Single-widget generation by prompt | Available now |
| One bounded continuity scenario across two runs | Demonstrated proof |
| Cortxt OS / Work, Evidence, Decisions | In development |
| Prompt-generated multi-widget composition ("Studio") | Direction |
Cortxt is a mandate/work-first operating layer: it’s built around what must persist across a run, not around any one agent framework or execution sandbox. That’s a different concern from an agent/workspace tool like OpenHands, or a runtime/security/sandbox layer like OpenShell/NemoClaw — Cortxt is designed to sit above changing execution systems, not to replace them.
We’re working directly with a small number of teams shaping the Work and Evidence surfaces before wider release. If durable, mandate-scoped AI work is a problem you have today, get in touch (temporary — opens a GitHub issue) ->
Start with what’s real today
No account, no hosted service to sign up for.
Clone the repository and install the CLI—the interface used today for mandate issuance, dispatch, and evidence collection.
macOS / Linux
# From a cloned repository
git clone https://github.com/rian010194/cortxt.git
cd cortxt
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ./agent-platform
cortxt --helpWindows PowerShell activates with .venv\Scripts\Activate.ps1 instead of the source line above — see the quick start for the full, verified sequence on both platforms.
This installs and verifies the CLI. Mandate issuance, dispatch, and evidence collection are real today but run through a configured GitHub repository, not a single global demo — the quick start walks through that full path.
Reference