JACO IDE: separating model output, tool execution and verified results in a local runtime

Hi everyone — I’m building JACO IDE, a local development runtime, and I’d like feedback on a problem that crosses model providers: how to separate a model’s proposed action, actual tool execution and evidence that the result is correct.

The current project connects an IDE, Aura chat, Nova execution and a local Workplace coordination service. JACO Editor MCP exposes the same workbench API used by the IDE, so external coding clients can work with the project through one interface.

Three concrete engineering choices:

  1. Conflict detection: file overwrites require the current content hash. If the file changes between read and write, the server rejects the stale write. Exact edits require a unique match.
  2. Bounded tools: project paths and supported Git actions have explicit boundaries. Commands/tests are handed to Nova through start_goal rather than exposing an unrestricted shell in the editor MCP server.
  3. Separate evidence: a connected client, an accepted goal and a verified result are different states. Startup and connectivity checks are reported as such, without treating them as proof that an arbitrary coding task passed.

For the current Windows Lite release candidate we checked startup from a separately extracted ZIP, runtime identity, Nova/Workplace readiness and a real MCP project read. The editor server was also exercised against the runtime for file operations, stale-write rejection and rejected paths/actions.

You bring your own model/provider; local options include Ollama and LM Studio. The longer-term aim is to retain repository knowledge and verified experience across tasks. That remains an architectural direction, not a claim that autonomous learning is complete.

Questions for other AI application builders:

  • What evidence do you require before labeling an agent’s task complete?
  • Which failure cases would you test first in an MCP workbench shared by multiple clients?
  • When trying a local coding runtime, what must work in the first ten minutes for you to keep using it?

The project, recorded run and Lite download are at https://jacocode.com/
Source: https://github.com/Rick84444/JACO-Autonomus-IDE

Windows is the tested package path; Node.js 22.5+ and full ZIP extraction are required. Lite does not include SEER or a managed public Cloud service. This post is a general runtime/project discussion, not a Gemini-specific integration benchmark. JACO is independent of Google.

I’m the project’s builder and would welcome practical technical feedback. Please redact secrets from any reports.