Multi-LLM Tool Descriptor Export

ABI exports provider-neutral tool descriptors for AI agent platforms via the unified tool_descriptors SSOT. The canonical command is:

abi export-tools --type metagenomic_plasmid --format openai --provider openai
abi export-tools --type metagenomic_plasmid --format openai --provider deepseek
abi export-tools --type metagenomic_plasmid --format anthropic
abi export-tools --type metagenomic_plasmid --format gemini

Supported provider profiles: OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Zhipu (GLM), Kimi (Moonshot), Qwen (Tongyi), MiniMax.

Default Tools

Default exports include:

  • abi_list_types

  • abi_plan

  • abi_check

  • abi_dry_run

  • abi_inspect

  • abi_report

  • abi_export_nextflow

  • abi_export_snakemake

  • abi_export_agent_context

  • abi_doctor_agent

  • abi_install_skills

  • abi_query

  • abi_validate_result

  • abi_autoplasm_validate_result (legacy compatibility alias)

abi_run is omitted unless --include-execution is passed.

Format Families

Three format families are supported:

  • OpenAI-compatible (--format openai): function descriptors for OpenAI, DeepSeek, Zhipu, Kimi, Qwen, MiniMax. Uses additionalProperties: false, strict: true (Responses API), and readOnlyHint (Apps SDK).

  • Anthropic (--format anthropic): tool_use descriptors for Claude.

  • Gemini (--format gemini): function_declarations for Google Gemini.

Agent Context

Agents can fetch compact operating context:

abi export-agent-context --type metagenomic_plasmid --format json
abi doctor-agent --type metagenomic_plasmid

The context lists safe call order, standard tables, important artifacts, error codes, and recovery rules.

Agent Skills (Claude Code)

ABI bundles SKILL.md files (one for each bioinformatics tool plus an abi_agent operating skill) inside the package at src/abi/skills/. Install them into Claude Code with:

abi install-skills         # → ~/.claude/skills/abi/
abi install-skills --force # overwrite existing files

Claude Code automatically loads all skills from ~/.claude/skills/ on each session start.

MCP Server

abi-mcp                    # safe profile: no execution or management tools
abi-mcp --profile discovery
abi-mcp --profile full     # add confirmation-gated abi_run
abi-mcp --profile management
python -m abi.mcp.server   # equivalent (if abi-mcp not available)

MCP visibility profiles are separate from the default descriptor export:

  • discovery exposes seven read-only discovery, preflight, and inspection tools;

  • safe exposes twelve tools, adding planning, dry-run, report, and workflow export while omitting execution and management;

  • full adds confirmation-gated abi_run;

  • management also exposes abi_install_skills and the legacy abi_autoplasm_validate_result alias.

Use full only when the session needs abi_run; execution still requires confirm_execution=true after explicit user approval.

Python API

import abi
print(abi.get_agent_guide())        # compact guide for system prompt injection
print(abi.list_plugins_summary())   # [(analysis_type, name, description), ...]