Brown Octopus

Brown Octopus manages the capability context exposed to an AI agent.

Given a user request, it identifies relevant capabilities, retrieves a bounded set of candidates, and maintains the active capability context for a conversation. It returns capability definitions to the host application. It does not call tools, own an LLM, manage credentials, or store conversation history.

Distribution: brown-octopus
Import:       brown_octopus
Runtime:      Octopus
Index facade: OctopusIndex

Install

Brown Octopus supports Python 3.12 and newer.

pip install brown-octopus

Or with uv:

uv add brown-octopus

Prepare the default runtime models explicitly:

brown-octopus setup-models

This prepares en_core_web_trf and Qwen/Qwen3-Embedding-0.6B. Octopus.initialize() never downloads models or discovers capabilities implicitly.

Check the installation:

brown-octopus doctor
brown-octopus inspect

Quick start

The normal lifecycle is:

CapabilitySource -> OctopusIndex.create() -> persisted index
                                              |
                                              v
                    Octopus.initialize() -> retrieve_result()

1. Configure a local MCP catalog

Create data/mcps.json:

{
  "items": [
    {
      "id": "outlook",
      "name": "Outlook Mail",
      "url": "https://example.com/outlook/mcp"
    }
  ]
}

The URL must be a reachable MCP server that the host is authorized to access.

2. Build the index

Create setup_index.py:

import asyncio
from pathlib import Path

from brown_octopus import OctopusIndex
from brown_octopus.sources import LocalMcpCatalogSource


async def main() -> None:
    index = OctopusIndex.from_sources(
        [LocalMcpCatalogSource(Path("data/mcps.json"))],
        index_path="data/indexes/default",
    )
    report = await index.create()
    print(f"Indexed {report.tool_count} capabilities")


if __name__ == "__main__":
    asyncio.run(main())

Run it once after the catalog is configured:

python setup_index.py

create() creates and publishes a snapshot only when the index does not already exist. It discovers capabilities, builds embeddings, and validates the snapshot atomically. To deliberately rebuild an existing index, use create(replace=True). It does not execute MCP tools.

3. Start the runtime

Create app.py:

import asyncio

from brown_octopus import Octopus


async def main() -> None:
    octopus = Octopus(index_path="data/indexes/default")
    await octopus.initialize()

    result = octopus.retrieve_result(
        "Reply to Tom's email",
        session_id="conversation-123",
    )

    print(f"Turn: {result.turn}")
    for capability in result.tools:
        print(capability["name"])


if __name__ == "__main__":
    asyncio.run(main())

Run it with python app.py. Pass result.tools to the agent or harness; the host decides how to bind and execute the definitions.

Index and runtime

OctopusIndex owns index lifecycle:

• create() creates the index only if it is missing;
• create(replace=True) deliberately rebuilds the index;
• update() synchronizes configured sources;
• add(source) incrementally adds a source to an existing index;
• add(source_a, source_b) batches several incremental additions;
• remove_source(source_id) removes a configured source and synchronizes;
• reset() removes the persisted index;
• runtime() returns the runtime facade.

Octopus owns runtime retrieval:

• initialize() loads models and the current snapshot;
• retrieve() returns the active capability context;
• retrieve_result() returns the typed result contract;
• session methods maintain conversation-specific capability state.

Sources are optional when opening an index:

index = OctopusIndex(index_path="data/indexes/new-index")
octopus = index.runtime()
await octopus.initialize()

result = octopus.retrieve_result("Reply to Tom's email")
assert result.tools == []
assert result.retrieved_tools == []
assert result.tool_ids == []

If the path has no snapshot, runtime initialization starts with an empty capability universe. Configure a source and call create() or update() when the host is ready to provision capabilities. Calling create() during every startup is safe: an existing snapshot is left unchanged.

Updating capabilities

After changing a configured source:

report = await index.update()
print(report.added)
print(report.changed)
print(report.removed)
print(report.failed_sources)

update() is a synchronization operation, not an append-only operation. It rediscovers the configured universe and publishes a complete replacement snapshot:

new capability                      -> added
changed metadata                    -> changed
unchanged capability                -> retained
authoritatively removed capability  -> removed
temporarily failed source           -> previous capabilities preserved

The host decides when updates happen. Brown Octopus has no refresh scheduler.

Add a source to an existing index:

from brown_octopus import McpServerSource

await index.add(
    McpServerSource(
        "https://example.com/word/mcp",
        source_id="word",
    )
)

add() requires an existing index. It discovers only the supplied source, embeds only new or changed capabilities, and preserves capabilities already in the snapshot. It does not perform a full source synchronization. To add a batch in one operation:

await index.add(
    McpServerSource("https://example.com/word/mcp", source_id="word"),
    McpServerSource("https://example.com/mail/mcp", source_id="mail"),
)

Call create() before the first add():

await index.create()
await index.add(McpServerSource("https://example.com/word/mcp"))

remove_capability() remains temporarily for older applications but is deprecated. Prefer changing the authoritative source or removing the source.

Retrieval results and filtering

result = octopus.retrieve_result(
    "Send Sarah an email",
    session_id="conversation-123",
)

The result provides:

result.retrieved_tools  current-turn selection
result.tools            final active context
result.tool_ids         stable IDs for result.tools
result.session_id       conversation identifier
result.turn             current session turn
result.metadata         timing and pipeline metadata

The simpler method is also available:

active_tools = octopus.retrieve(
    "Search the web for the latest NVIDIA news",
    session_id="conversation-123",
)

Restrict a turn before ranking with allowed_mcp_urls:

result = octopus.retrieve_result(
    "Reply to an email",
    session_id="conversation-123",
    allowed_mcp_urls=["https://example.com/outlook/mcp"],
)

Sessions

One engine can serve many conversations:

first = octopus.retrieve_result("Search the web", session_id="conversation-a")
second = octopus.retrieve_result("Reply to Tom", session_id="conversation-b")

Each session has its own turn counter and active capability state. A session ID represents a conversation or workflow identity, not a user identity.

snapshot = octopus.get_session("conversation-a")
octopus.reset_session("conversation-a")
octopus.delete_session("conversation-b")

The default store is process-local. A persistent application can inject a custom atomic SessionStore; see docs/production.md.

Capability sources

Built-in adapters include:

Convenience constructor:

index = OctopusIndex.from_mcp_server(
    "https://example.com/outlook/mcp",
    source_id="outlook",
    mcp_headers={"Authorization": "Bearer <short-lived-token>"},
    index_path="data/indexes/outlook",
)

Detailed source contracts, pagination, authentication, and field mapping are documented in docs/production.md.

Embedding providers

The default provider is local Qwen. A compatible local model can be supplied:

from brown_octopus import LocalEmbeddingProvider, OctopusIndex

provider = LocalEmbeddingProvider(
    model_name="sentence-transformers/all-MiniLM-L6-v2",
)
index = OctopusIndex.from_sources(
    [source],
    index_path="data/indexes/minilm",
    embedding_provider=provider,
)
await index.create()

An HTTP provider can target a remote or self-hosted service:

from brown_octopus import HttpEmbeddingProvider, OctopusIndex

provider = HttpEmbeddingProvider(
    url="http://localhost:11434/api/embed",
    model="qwen3-embedding:0.6b",
)
index = OctopusIndex.from_sources(
    [source],
    index_path="data/indexes/http-embeddings",
    embedding_provider=provider,
)
await index.create()

Use the same provider when loading that index. Provider dimensions are checked against index metadata.

CLI

brown-octopus --help
brown-octopus --version
brown-octopus version
brown-octopus setup-models
brown-octopus doctor
brown-octopus inspect
brown-octopus inspect --json

Help and version do not load models. doctor checks package, models, and the configured index. inspect reports configuration and index metadata.

Runtime context limits

The default active-context policy is:

TTL:       8 turns
Active cap: 30 capabilities

These defaults preserve the validated behavior, but applications can tune them when their context budget requires it:

from brown_octopus import OctopusIndex

index = OctopusIndex(index_path="data/indexes/default")
await index.create()
octopus = index.runtime(
    capability_ttl=3,
    active_cap=20,
)
await octopus.initialize()

capability_ttl controls how many later turns a retrieved capability can remain active without being retrieved again. active_cap limits the total number of active capabilities in a session. These settings affect session context management, not candidate ranking or the V3 selection algorithm.

The same values can be configured through environment variables:

OCTOPUS_CAPABILITY_TTL=3
OCTOPUS_ACTIVE_CAP=20

Architecture boundary

Brown Octopus owns capability discovery adapters, indexing, intent analysis, retrieval, selection, active capability context, and capability session state.

The host owns the LLM, conversation history, tool binding and execution, credentials, authentication, authorization, user identity, update timing, and session lifecycle policy.

Returned MCP metadata such as mcp_url and tool_name describes execution context. Brown Octopus does not execute the capability.

Retrieval behavior

The default V3 pipeline is unchanged:

user request -> deterministic spaCy intent analysis
             -> capability-oriented representation
             -> dense embedding retrieval
             -> Min-4 + Bounded Max Gap
             -> active capability context
             -> TTL 8 and active cap 30

The strategy is recall-oriented within bounded limits. It may expose extra plausible capabilities rather than aggressively pruning a capability that may be required. It does not guarantee perfect recall.

Development and research

uv sync
uv run brown-octopus setup-models
uv run pytest -q -m "not external" tests
uv build

Product code, tests, examples, and research/evaluation history are kept separate. Evaluation material under evals/ and docs/research/ is retained for reproducibility and is not required at runtime.

See docs/production.md for the complete integration reference.

License

See LICENSE.