TarsTARS

Local Extensibility

Extend Tars with reviewed memory, instructions, and executable MCP tools.

Tars can write files in its workspace when an enabled tool grants that access. This makes it possible to draft facts, skills, prompts, or extension code, but does not make unsupervised self-modification safe.

Levels of change

  1. Durable memory stores an explicit fact or note. This changes context, not executable code.
  2. System instructions and skills change future model guidance. Review them for hidden commands, credentials, or conflicting policy.
  3. MCP extensions are executable subprocesses. Treat installation as a code-deployment event.

Review workflow

For instructions or code generated by the model:

  1. define the requested behavior and allowed authority;
  2. stage the change outside the active installed path;
  3. review code, dependencies, manifest, secrets, and failure paths;
  4. run focused tests with read-only inputs;
  5. add an explicit enablement entry only after review;
  6. ask the operator to restart and verify startup logs;
  7. retain a rollback copy.

Tars does not sign, sandbox, or independently verify generated extensions. Avoid allowing a model to install dependencies or enable its own executable code without operator review.