Splice

Zero-friction context sync between Kiro, Cursor, and Claude Code.

Splice is an invisible meta-tooling layer for any codebase. Install once, then forget about it. It silently tracks context as you work, syncs it across IDEs and machines via git, and ensures every AI assistant has the full picture — without you ever copy-pasting an error log again.

Three orchestration modes: mock (offline) · Ollama (local, privacy-first) · Bedrock (cloud).

Install (one time)

pip install splice-cli
cd your-project
splice setup

That's it. Splice is now active. No commands to remember, no workflow to learn.

How it works

flowchart TD
    subgraph "You just work normally"
        W[Write code in any IDE]
        W --> P[Every prompt silently tracked]
        P --> C[Commit your code]
    end

    subgraph "Splice handles the rest"
        C -->|pre-commit hook| B[Bundle updated & committed]
        B -->|git push/pull| S[Synced to any machine]
        S -->|SessionStart hook| I[Next IDE auto-loads bundle]
    end

Why this exists

If you use Kiro, Cursor, and/or Claude Code on the same project, you probably:

• Paste the same error log in two IDEs
• Lose session continuity when switching tools
• Burn tokens sending irrelevant files to the model
• Context-switch between machines and lose everything

Splice fixes all of this without changing your workflow.

Privacy-first: Ollama (recommended)

Your code and context never leave your machine:

brew install ollama
ollama pull llama3.2

Edit .splice/config.json (created by splice setup):

{ "orchestrator": "ollama", "ollama": { "model": "llama3.2" } }

Kiro/Cursor/Claude Code still use their own cloud models for chat. Splice controls only the orchestration (routing + context reduction).

Optional power-user commands

Splice works automatically, but these are available if you want manual control:

splice status                    # show active session
splice doctor                    # verify orchestrator health
splice capture-error             # pipe test output: npm test 2>&1 | splice capture-error
splice upgrade                   # upgrade to latest version + update hooks
splice uninstall                 # remove all Splice hooks from current project
/harness <task>                  # explicitly request bundle in any IDE prompt

What gets committed

Splice adds one file to your repo:

.splice/active-bundle.md    ← committed (synced via git)
.splice/session.json        ← gitignored (local only)
.splice/last-error.log      ← gitignored (local only)

The bundle is a markdown summary of your current context: errors, relevant files, diffs, and the recommended agent/workflow. It's small, human-readable, and useful for code review too.

Configure (optional)

.splice/config.json (created automatically by splice setup):

{
  "orchestrator": "mock",
  "captureGitDiff": true,
  "includePaths": ["src/**/*.ts"],
  "errorSources": [".splice/last-error.log"],
  "ollama": { "model": "llama3.2" },
  "bedrock": { "modelId": "amazon.nova-lite-v1:0" }
}

Supported IDEs

How context flows

Machine A (Kiro)          Git repo              Machine B (Cursor)
─────────────────         ────────              ──────────────────
Work normally        →    commit includes       →    Open Cursor
Context tracked           active-bundle.md           Bundle auto-injected
silently                  push to remote              on SessionStart
                                                     Full context ready

Requirements

• Python 3.10+
• Git
• Kiro, Cursor, and/or Claude Code
• Ollama (optional, for local LLM orchestration)

Docs

Contributing

PRs welcome — especially IDE adapters (Windsurf, Zed) and Ollama model presets.

License

MIT — see LICENSE.

Author

Created by Hazel Saenz — LATAM Developer Advocate at AWS | Former AWS Serverless Hero.