Drive the agent tools from Claude Code¶
New in 0.2
astro-weights mcp RUN_DIR serves the agent tools over stdio MCP. Add it to Claude Code as a local server and Claude Code becomes the editor: it calls exec_python, forward_pass, status and finish on your model.
You need the mcp extra, plus hf for forward_pass:
1. Make a run directory¶
The server reads RUN_DIR/run.json. model_dir is read relative to RUN_DIR. Every key is listed in Reference: CLI. Tools, result keys and refusals: Reference: MCP server.
{
"seed": 0,
"editor": "claude-code",
"model": null,
"limits": {"forward_passes": 20, "code_runs": 50, "tokens": 0, "dollars": 0},
"model_dir": "/path/to/model"
}
Save it as /path/to/run/run.json. Use one run directory per run: a restart with a changed seed, editor, model or limits is refused.
2. Add the server¶
Everything after -- is the server command. Use an absolute RUN_DIR; Claude Code may start the server from another directory. --scope project writes the server to .mcp.json instead (Claude Code MCP docs).
The same as a .mcp.json entry:
{
"mcpServers": {
"astro-weights": {
"command": "astro-weights",
"args": ["mcp", "/path/to/run"]
}
}
}
3. Run¶
Start claude and ask it to edit the model, for example "Use the astro-weights tools to find an edit that shortens answers; call finish when done." The tools are named mcp__astro-weights__exec_python and so on; Claude Code asks before each call unless you allow them.
exec_python and forward_pass return a handle at once. Claude polls status(handle) until the job is done or failed. Read arrays in exec_python with forward_result(handle, key).
Limits¶
- The server counts
code_runsandforward_passesand refuses work past them. Claude Code's own tokens and dollars are not reported to the server, sotokensanddollarslimits do not cap a Claude Code session. Your own Claude Code limits and your provider's terms apply. - Claude Code also sees the three editor tools
report_usage,check_spendandreport_refusal. The built-in editors hide them from their model; Claude Code does not. - No OS sandbox (Windows, or a blocked bwrap):
exec_pythonreturns an error.astro-weights mcp RUN_DIR --unsafe-kernelruns the code unsandboxed; see Explanation: agent budgets and sandbox.
Tested once: a headless claude -p run on a tiny random model (0.2.0a1, macOS) used exec_python, forward_pass, status and finish, and the run finished with an exported adapter.