tako-ai-core 0.51.2

Rust-core, Python-facade framework for enterprise agentic systems.

tako 蛸

Rust-core, Python-facade framework for enterprise agentic systems.

Many arms, one mind.

tako is an open-source framework for building production agentic systems. It gives you vendor-neutral provider abstractions, a Rust orchestration core that keeps Python's GIL out of the hot path, MCP tool connectivity, and the governance plumbing (OTel tracing, OPA policy, PII redaction, budgets, circuit breakers) you actually need at scale — all with a Pythonic, dual sync/async API that ships as one pip install.

Inspiration & credit

tako is an open-source generalisation of three patterns Sakana AI published, plus AB-MCTS tree search:

1. Trinity-style learned routing — a small model selects which provider/role handles each step. Xu et al., "TRINITY: An Evolved LLM Coordinator," arXiv:2512.04695.
2. Conductor-style natural-language orchestration — a coordinator agent decomposes tasks and dispatches workers. Nielsen et al., "Learning to Orchestrate Agents in Natural Language with the Conductor," arXiv:2512.04388.
3. Self-recursive test-time scaling — bounded recursion in which an agent reads its own output and decides whether to spin up corrective workflows. See Sakana AI's Fugu Beta blog post.
4. AB-MCTS — Adaptive Branching Monte Carlo Tree Search. Inoue et al., arXiv:2503.04412; reference implementation by Sakana AI as TreeQuest (Apache-2.0).

tako is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Sakana AI or any model provider. The cited papers are credited as inspiration for the underlying patterns; the implementation is the work of the tako contributors. The name tako ("octopus") complements Sakana AI's "Fugu" (pufferfish) as a tribute.

Install

pip install tako-ai-core

The Python import name is still tako (import tako); the PyPI distribution is tako-ai-core because the bare tako slot was already taken by an unrelated 2011-era project.

No Rust toolchain required at install time — wheels are prebuilt for manylinux, musllinux, macOS universal2, and Windows x64/arm64.

Quickstart

import asyncio
import tako

client = tako.Client(
    providers=[
        tako.providers.Anthropic(model="claude-opus-4-7"),
        tako.providers.OpenAI(model="gpt-5"),
    ],
    mcp_servers=[
        tako.mcp.Stdio(command=["npx", "-y", "@modelcontextprotocol/server-everything"]),
    ],
    tracing=tako.tracing.Otlp(endpoint="http://otel-collector:4317"),
    budget=tako.Budget(max_usd_per_request=5.0, max_usd_per_day=500.0),
)

orch = tako.orchestrator.SingleAgent(
    provider="anthropic:claude-opus-4-7",
    max_steps=10,
)

async def main():
    result = await orch.run("What's the weather in Tokyo? Use a tool.")
    print(result.text)

asyncio.run(main())

# Synchronous sibling:
result = orch.run_sync("Quick question: ...")

Capabilities (current release)

Area	What's available today
Providers	OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex (Gemini), Mistral, Ollama, plus an http-generic template adapter and a PythonProvider. All seven SDK-backed providers handle outbound vision content (inline + URL-source); Bedrock + Ollama use opt-in tako-side URL pre-fetch with the full SSRF mitigation stack (private-IP blocklist + DNS-rebind defence + per-host / wildcard / CIDR allowlist).
Orchestrators	SingleAgent, Conductor, Trinity (rule-based or ONNX-backed router), SelfCaller (bounded recursion), AbMcts (Adaptive Branching MCTS with verifiers + router-driven branch expansion). All stream natively via OrchEvent.
Streaming	Native provider.stream(...) on every SDK-backed provider; per-delta OrchEvent::VerifierScore on Trinity, Conductor, AbMcts (bounded mpsc::channel(64) for backpressure); streaming-aware ConfidenceGuard (RuleBasedGuard + opt-in LlmJudgeGuard per-N-delta).
MCP	Stdio, Streamable HTTP (with notifications() SSE + Mcp-Session-Id lifecycle), WebSocket, gRPC (with mTLS).
OpenAI-compat HTTP server (tako-compat)	Drop-in /v1/chat/completions with tako.* SSE extensions (tako.verifier_score, tako.recursion, tako.tool_call_*). Pluggable AuthResolver: static, JWT, OIDC, Vault, plus a composite ChainedAuthResolver with opt-in transport / infrastructure-error fail-fast. OIDC introspection ships every RFC 7662 / 8414 / 8705 auth method including mTLS with explicit cert/key rotation.
Governance	OPA / Rego policy (Allow / Deny / RedactMessages / ForceModel / RequireApproval), PII / DLP redaction, OTel tracing with tako.* + gen_ai.* semconv, in-memory + Redis budgets, circuit breakers, rate limits.
Sigstore	Tool-catalogue keyed + keyless verification with operator-pinned trust roots, Rekor SET + inclusion-proof + checkpoint freshness anchor (in-memory / on-disk JSON / Redis-backed StateStore), cosign protobuf-bundle.
Reliability	Cascade fallback, governor rate-limit, failsafe circuit breaker, exponential-jitter retry.
API surface	Sync + async dual API; mypy-strict types; full Pydantic v2 facade.

For the chronological ledger of which capability landed in which phase (33 phases × 17 rows), see docs/feature_matrix.md or the Feature matrix page on the docs site.

Project history

The project ships in numbered phases. The rolling per-phase index lives in PLAN.md, individual plan documents under plans/, and the per-version release notes in CHANGELOG.md.

For a high-level summary of what shipped when, the docs site has a Feature matrix page (source).

Community

• Issues: https://github.com/nyankobu010/tako-ai-core/issues
• Discussions: https://github.com/nyankobu010/tako-ai-core/discussions
• Security: see SECURITY.md — please use GitHub Private Vulnerability Reporting rather than opening a public issue.
• Good first issues: https://github.com/nyankobu010/tako-ai-core/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22

License

Apache-2.0 — see LICENSE and NOTICE.