thisispandora-agent

thisispandora-agent is the standalone Python SDK for Pandora's agent-facing contract and MCP execution surfaces.

It is intended to work as a normal installed Python package:

• it ships its own generated contract artifacts inside pandora_agent/generated
• it can execute against local stdio MCP via pandora mcp
• it can execute against operator-hosted remote MCP HTTP via pandora mcp http
• it exposes package-local helpers for the vendored manifest, contract registry, command descriptors, and MCP tool definitions
• it exposes first-class bootstrap helpers so cold agents can start from Pandora's canonical bootstrap contract

What this package is for

Use thisispandora-agent when Python code needs to:

• inspect the shipped Pandora command and tool catalog without shelling out
• inspect policy scopes and signer-profile metadata before choosing tools
• connect to a local Pandora process over stdio MCP
• connect to a remote Pandora MCP HTTP gateway with a bearer token

Prerequisites

The Python package does not bundle the Pandora CLI runtime itself.

• For local stdio execution, the pandora CLI must be installed and available on PATH, or you must pass an explicit command=....
• For remote execution, an operator must already be running pandora mcp http ... and provide the /mcp URL plus a bearer token if auth is enabled.

Installation

Current validated install paths:

Preferred for external consumers:

pip install /path/to/downloaded/pandora_agent-<version>-py3-none-any.whl

Maintainer and repository-checkout flows:

pip install ./sdk/python

Or from a locally built release artifact produced by the repository release flow:

pip install dist/release/sdk/python/*.whl

Use the signed GitHub release wheel or sdist when working outside the repository. Use the repo path only when you intentionally want an in-tree checkout. Public PyPI publication is not claimed by this README until a release explicitly says so.

Vendored equivalent inside the Pandora CLI package:

PYTHONPATH=/path/to/pandora-cli-skills/sdk/python python

Quickstart

Local stdio MCP

Cold agents should start with bootstrap, not with low-level capabilities or schema calls.

from pandora_agent import create_local_pandora_agent_client

with create_local_pandora_agent_client(command="pandora") as client:
    bootstrap = client.get_bootstrap()
    print(len(bootstrap["canonicalTools"]), bootstrap["recommendedBootstrapFlow"][0])

Remote MCP HTTP

from pandora_agent import PandoraSdkError, PandoraToolCallError, create_remote_pandora_agent_client

with create_remote_pandora_agent_client(
    url="http://127.0.0.1:8787/mcp",
    auth_token="replace-me",
) as client:
    try:
        bootstrap = client.get_bootstrap()
        print(bootstrap["canonicalTools"][0])
    except PandoraToolCallError as error:
        print(error.code, error.details)
    except PandoraSdkError as error:
        print(error.code, error.details)

Package-local contract inspection

from pandora_agent import (
    inspect_generated_command_policy,
    load_generated_command_descriptors,
    load_generated_manifest,
)

manifest = load_generated_manifest()
trade_descriptor = load_generated_command_descriptors()["trade"]
trade_policy = inspect_generated_command_policy("trade")

print(manifest["packageVersion"])
print(trade_descriptor["canonicalTool"])
print(trade_policy.policy_scopes)

Main API surface

Client constructors

• create_local_pandora_agent_client(...)
• create_remote_pandora_agent_client(...)
• PandoraAgentClient(...)

PandoraAgentClient supports context-manager usage, so installed consumers can write with ... as client: and avoid manual teardown.

Generated artifact helpers

• load_generated_manifest()
• load_generated_contract_registry()
• load_generated_capabilities()
• load_generated_command_descriptors()
• load_generated_mcp_tool_definitions()
• get_generated_artifact_path(...)
• list_generated_artifact_paths()

Policy/profile helpers

• load_generated_policy_profiles()
• inspect_generated_command_policy(command_name)
• PandoraAgentClient.get_policy_profiles()
• PandoraAgentClient.inspect_command_policy(command_name)

Error semantics

The SDK raises:

• PandoraSdkError for SDK, transport, protocol, process, or HTTP failures
• PandoraToolCallError for Pandora tool failure envelopes

Tool-call normalization is intentionally specific:

• when Pandora returns a structured tool error with its own stable code, PandoraToolCallError.code preserves that Pandora code directly
• examples: FORBIDDEN, UNKNOWN_TOOL, MCP_INVALID_ARGUMENTS
• the generic wrapper code PANDORA_SDK_TOOL_ERROR is only used when a tool failure envelope does not include a stable Pandora error code

PandoraSdkError.to_dict() and PandoraToolCallError.to_dict() are safe to log or serialize for debugging. PandoraToolCallError also keeps the normalized envelope and raw MCP result payload when available.

Transport notes

Local stdio

• starts pandora mcp
• keeps one subprocess per client instance
• requires connect() before use unless you use the context-manager form

Remote HTTP

• talks to the streamable HTTP MCP endpoint at /mcp
• supports bearer-token auth via auth_token=...
• preserves gateway-provided MCP session headers when the server emits them
• surfaces HTTP and gateway-side JSON errors as PandoraSdkError

Examples

Repository examples are kept under:

• sdk/python/examples/local_stdio.py
• sdk/python/examples/remote_http.py

They are source-tree examples, not installed package files in the wheel.

Current boundaries

This package is intentionally standalone at runtime, but the repository and the Pandora CLI package also vendor matching copies and share some integration boundaries:

• the vendored generated JSON artifacts are still produced by the shared contract generator in the main repo
• package versioning and release or publish automation outside sdk/python/** are still controlled by shared repo release work
• the Pandora CLI package also carries a vendored copy under sdk/python for in-tree consumers and parity checks

That means the installed package is self-contained at runtime, but artifact refresh and coordinated release automation still depend on the repo-level generator and release lanes.