Plan required before implementation.
Desktop control plane / AI-assisted software work
Provable AI delivery
Twindem turns AI-assisted work into governed delivery: every idea starts on the board, every phase crosses a gate, and every outcome keeps a trail.
- Board source of truthOK
- Typed idea intakeOK
- Dual-agent reviewOK
- Human gatesOK
- Release evidenceOK
Board
source
Agent 2 has open findings.
Waiting for human gate.
Agent loop
running
A1
Author
Drafts plan, edits code, records implementation notes.
A2
Reviewer
Challenges the work, opens findings, triggers the fix loop.
Human gate
Approve UAT
Evidence
kept
- 09:12 Board task selected
- 09:24 Refinement approved
- 10:06 Review finding resolved
- 10:18 UAT runbook attached
01 / System
Not another chat window. A delivery system around AI-assisted work.
Board first
Ideas become tracked work before the agents start. Type, status, scope, and evidence stay attached to the task.
Two roles
One agent authors and implements. The second reviews, challenges, and keeps the loop honest.
Human gates
The team decides when a plan becomes implementation, when code reaches UAT, and when production is allowed.
Release memory
Runbooks, phase changes, review notes, and evidence are preserved instead of buried in terminal scrollback.
02 / Workflow templates
Not every idea should become a coding task.
Twindem lets teams classify work as Feature, Bug, Architecture, Research, or Runbook, then applies the right workflow template for each. Same board. Same human gates. Different evidence requirements for different kinds of work.
Features and bugs need implementation. Architecture, research, and runbooks need decisions, evidence, and validation. Twindem understands the difference, so AI work does not skip straight from idea to code.
Technical plan, acceptance criteria, implementation, code review, UAT, and release.
Reproduction, root cause, fix, regression review, verification, and confirmed resolution.
Options, ADR draft, challenge review, stakeholder approval, and accepted decision record.
Questions, investigation, critique, decision checkpoint, and accepted recommendation.
Scope, procedure draft, safety review, dry run, and approved operating procedure.
Type
Planning
In progress
Review
UAT
Done
Feature
Technical plan + acceptance criteria
Implementation
Agent 2 code review
User testing / validation
Merged or released
Bug
Reproduction + root cause
Fix
Regression review
Bug fix verification
Confirmed fixed
Architecture
Options + ADR draft
Decision document / optional proof of concept
Challenge review
Human / stakeholder approval
ADR accepted
Research
Research questions + scope
Investigation / comparison
Critique findings
Decision checkpoint
Recommendation accepted
Runbook
Scope + preconditions
Procedure drafting
Safety / reliability review
Dry-run / validation
Runbook approved
Consistent statuses, different meaning.
Every idea can move through inbox, planning, in progress, review, UAT, and done. The template defines what each phase means, what agents should produce, and what evidence must exist before approval.
03 / Workflow
Every phase is a tandem, not a handoff.
Capture, refine, build, review, and validate are not just columns on a board. Agent 1 produces the work product, Agent 2 challenges it, and the human gate decides when the evidence is enough to move forward.
Automation is intentionally scoped: agents can loop inside the current phase, but moving to the next phase is a manual user action.
Inbox
Idea intake
Choose Feature, Bug, Architecture, Research, or Runbook. The type sets the evidence expected later.
Planning
Refine
Agent 1 drafts the plan, scope, ADR outline, research questions, or runbook preconditions.
A1
drafts
A2
challenges
H
gates
In progress
Produce
For features and bugs this means code. For architecture, research, and runbooks it means the right work product.
A1
creates
A2
flags gaps
A1
fixes
Review
Challenge
Agent 2 performs an independent review against the idea type, not just a generic code check.
A2
reviews
A1
resolves
A2
verifies
UAT
Validate
Human validation confirms the right thing was produced: release candidate, ADR, recommendation, or procedure.
A1
presents
H
validates
✓
records
Done
Evidence accepted
The board can move forward only after the required output and review evidence exist.
04 / Evidence
When someone asks what happened, the answer is not “check the chat”.
Review loop
Agent 2 review output and Agent 1 fixes are tied to the delivery state, not treated as disposable conversation.
Runbooks
UAT and production steps live with the workspace, so releases follow the same instructions each time.
Phase history
Planning, implementation, review, UAT, and production transitions can be retained as workflow events.
05 / Agent stack
Use the CLIs your team already understands.
Twindem orchestrates local agent terminals instead of hiding work in a hosted black box. Humans can still talk directly to the agents when a task needs judgment.
Agents
Claude Code, Codex, shell fallback
Board providers
GitHub Projects and Jira
Auth
CLI login or API keys
Runtime
macOS Electron desktop app
06 / Board setup
Bring your own board — and your own keys.
Jira and GitHub Projects, side by side
Pick a board provider per project. A Jira-tracked project and a GitHub-tracked one live in the same Twindem, with no conflict.
Connect once, pick or create
Authenticate Jira with your email and API token, then choose an existing project from the list — or create a new one without leaving the app. GitHub works through your local gh session.
Your credentials stay yours
Your Jira API token is encrypted locally in the OS keychain — never stored in plaintext, never sent anywhere but Atlassian. GitHub uses your local CLI session; Twindem stores no GitHub passwords or tokens.
Local-first, single-user
Everything lives on your machine. No Twindem cloud, no sync, nothing to opt out of.
Board and code, kept separate
The board is where tasks live; a separate Code/Repo setting controls where agents are allowed to work. They are never confused.
07 / Cost & context
Two agents should not cost twice.
Running an author and a reviewer means every duplicated byte of context is paid twice. Twindem actively minimizes that, without compacting away the reasoning that makes delivery decisions trustworthy.
Built into the free app.
Governed context
Each agent receives a controlled, phase-aware brief scoped to its role. The author and the reviewer never get the same context, and agents stop re-fetching content Twindem already holds.
Compact evidence
Readable records of what was decided, produced, challenged, and approved — so humans inspect the task without re-reading terminal scrollback.
Cost visibility
Estimated context volume per task, phase, agent, and review round, with warnings when a loop runs hot. Honest numbers: estimates are labeled as estimates, never as billed tokens.
Token economy
Review rounds scoped to specific findings instead of full re-reviews, and agent restarts with a compact handoff that keeps decisions, constraints, and open findings — and drops the noise.
Loop guardrails
Duplicate signals are ignored, repeated terminal redraw is compacted, and hot review loops stop for human confirmation instead of spending tokens indefinitely.
08 / Today and next
What is in the app today, and where it is going.
In the free app today
Open source- Board-driven workflow on GitHub Projects or Jira
- Claude Code and Codex side by side, in real terminals
- Review and fix loop between the two agents inside the active phase
- Manual human gates for every phase transition
- Typed intake: feature, bug, architecture, research, runbook
- Project setup wizard for first-run and new projects
- Multiple projects with a workspace switcher
- Status mapping that adapts to existing board layouts
- Evidence, phase history, and release runbooks, stored locally
- Governed, role-scoped context for each agent
- Cost visibility per task, phase, agent, and review round
- Token economy: delta reviews, duplicate-signal guardrails, compact terminal output, and compact restarts
Commercial direction
Later- Security gate: automated security checks on every delivery, validating changes against industry standards (OWASP, CWE, dependency and secret scanning) before a release is approved
- Confluence integration: write and publish pages
- Built-in mockup designer for UI proposals
- Windows and Linux builds
- Homebrew install
- Linear board provider
- Standardized project provisioning with org workflow templates
- Team accounts and a shared audit timeline
- Team-level cost analytics and exact API usage reporting in headless/API mode
- SSO and organization-level controls
Built for companies that need accountability at scale, priced per seat.
09 / Screenshots
See Twindem in action.
10 / FAQ
Frequently asked questions.
What is Twindem?
A macOS desktop app that orchestrates two CLI AI coding agents through a board-driven delivery workflow. One agent authors and implements, the second reviews and challenges the work, and a human approves every phase transition. Evidence, phase history, and release runbooks are kept as an audit trail.
Which AI coding agents does Twindem support?
Twindem orchestrates Claude Code (Anthropic) and Codex (OpenAI) running locally in real terminals, side by side. You sign in with your own subscriptions or API keys, and you can talk to each agent directly in its terminal.
Is Twindem free and open source?
Yes. Twindem Community Edition is free and open source under the GNU AGPL-3.0 copyleft license. The source repository is Twindem-AI/twindem-community. A separate commercial license is available for use that cannot meet the AGPL obligations, and a commercial tier for companies is planned later, with Linear, team accounts, hosted sync, SSO, and organization controls.
Does my code leave my machine?
Twindem itself runs entirely on your machine and sends nothing to a hosted service. Sessions, evidence, and workflow history are stored locally. The AI agents run locally too, talking to their model providers with your own accounts or API keys, which are stored in the OS keychain.
How does Twindem store my API keys and tokens?
Every secret you enter — your Anthropic (Claude) and OpenAI (Codex) API keys and your Jira API token — is encrypted on your machine with your operating system's keychain via Electron safeStorage (macOS Keychain, Windows DPAPI, Linux secret service). The config file stores only a reference to each secret, never its value, and nothing is written to disk in plaintext.
Secrets stay local to the desktop app. The app has no Twindem cloud backend: your Anthropic/OpenAI key is passed as an environment variable to the local agent CLI when it runs, and your Jira API token is used only for requests to your own Jira site.
11 / Download
macOS builds live in GitHub Releases.
Public downloads are distributed from the Twindem Community Edition release page. Apple Silicon and Intel builds are published as separate DMG assets.
Open download page12 / License
AGPL-3.0 for the open-source core.
Twindem Community Edition is released under the GNU AGPL-3.0 copyleft license. Copyright and product rights are retained by the project owner, and a separate commercial license is available for use that cannot meet the AGPL obligations.
Read license details13 / Teams
Building Twindem for teams. Want early access?
Twindem Community Edition is free and open source. A commercial tier for teams — security gate, Linear, team accounts, a shared audit timeline, and SSO — is in the works. If your team wants it, tell us and we'll keep you in the loop.
Register interest
twindem.ai