How to share project context between Claude Code, Codex, and Cursor
Coding agents have saved conversations and instruction files, but those do not automatically become a shared working brief across tools. Switching agents can mean rebuilding the latest state by hand, losing the reason behind a constraint, or treating an unfinished experiment as an agreed decision.
A simple way to run it in Hark
- 01Connect every client to one hub
Point Claude Code, Codex, Cursor, and other supported MCP clients at the same Hark workshop. Choose the intended project and approve access for each connection. Keep access keys out of chat; use the client’s connection flow or a single-use pairing code. Workflow source
- 02Read the brief at session start
Call start_session first, then get_agent_brief before doing work. Check the project, Current state, accepted decisions, blockers, and next action. Ask the incoming agent to summarise its task so a mistaken assumption is visible before code changes. Workflow source
- 03Write decisions as they happen
Record durable constraints and rationale in the journal instead of leaving them in one chat transcript. Update living state as work changes. Agent proposals follow project trust settings, but decisions, non-goals, phase changes, and conflicts with accepted decisions still require review. Workflow source
- 04Save the handoff
Call end_session with what changed, what comes next, and what needs caution. Include tests actually run and anything still unverified. Have the following client read the handoff and use confirm_handoff to mark it as picked up. Workflow source
- 05Use the guardrails and repo export
Hark sessions are bound to one project; another requires an explicit switch. Quiet sessions get a draft handoff from recorded activity, not a recovered transcript. Optional repo export commits accepted content into .hark/HANDOFF.md, .hark/DECISIONS.md, and an AGENTS.md block. Workflow source
Resume a conversation or switch tools deliberately
If you use Claude Code across multiple projects, saved local sessions already help you return to earlier work. Claude Code saves CLI conversations locally as you work. claude --continue reopens the most recent session in the current directory; claude --resume opens a picker. Those are useful when the same conversation is still the right starting point, but they do not transfer its working context into Codex or Cursor. See the Claude Code session guide.
Context management is a separate choice. /compact replaces conversation history with a summary; /clear starts a fresh conversation while leaving the previous local session available to resume. Neither command publishes a cross-tool project record. Hark gives the next agent a deliberate account of the current work, rather than asking it to reconstruct which parts of an old conversation still apply.
Keep stable rules beside code and living state in Hark
Build commands and coding conventions belong in repository instructions. Claude Code supports persistent instructions and machine-local auto memory; Codex AGENTS.md and Cursor Rules also provide repository guidance. Hark does not need to replace those mechanisms. Keep stable rules in the repo and changing progress, blockers, decisions, and next actions in the project record.
Hark’s Inventory, Health, Work, Authority, and Decisions blocks supply the same underlying record to briefs and handoffs. Linked repository instructions remain authoritative, and Hark mirrors them rather than maintaining competing edits. Accepted does not mean independently verified: Agreed, Observed, and Proposed retain different meanings. See how Hark works.
For automatic briefing and compaction protection in Claude Code, follow the Hark mod connect guide. The mod is early access and requires Claude Code 2.1.287 or newer. Other clients use their connection guides and explicit session tools; the Claude Code mod’s automation is not universal.
Client handoff: an AI-built app checklist
A useful client handoff answers practical questions: where the work lives, what works, what remains uncertain, which constraints must survive, and who owns the next action. Use that checklist to review the project record before sharing. Include evidence for completed work, distinguish missing checks from failures, and describe access locations without storing credentials.
Sharing a Claude Code web session serves a different purpose. Web sessions have visibility settings and shareable links; local CLI transcripts are files, and /export produces readable text. A web session link is not a shared brief for every coding tool. Review sensitive content before sharing, as the cloud session sharing documentation advises. Hark’s recipient-first handoff adds a gap check, versioned handoffs, and a place for recipient questions.
Hark’s public project page
Starter builds are coming. Until then, use Hark’s own documented build as the example structure for this setup.
View the public project page placeholder →Common questions
- Can I use this setup for an existing AI-assisted project?
Yes. Import a conversation or brain dump, review the proposed Current state, and connect the repository when ready. You do not have to start from scratch. Product source Plan source
- Do all collaborators and agents see the same information?
They read the same underlying project record, subject to access permissions and section controls. Guests can read enabled sections and comment, not edit the project. Product source Plan source
- Does Hark copy my whole Claude Code conversation?
No. The Claude Code mod sends a handoff summary, not the transcript, prompts, file contents, or diffs. Closing-note redaction is best effort, so avoid putting secrets in prose. Product source Plan source
- Can I try this without adding a card?
Yes. Hark starts with a 14-day card-free trial. If it ends before you choose a plan, your workshop becomes read-only and your work remains available. Product source Plan source