Agents that survive, resume, and answer for themselves.
An in-process runtime you can embed anywhere. Git-backed context gives you crash-resume, rewind, and an audit trail for free. Model-agnostic by construction — no upstream vendor SDKs.
The API
Two methods to learn
send() and close(). The context directory is a real git repository — run against it again and the conversation picks up where it left off. The audit record is not a separate artifact; it is the git tree.
In practice
The properties you'd otherwise build by hand
Crash, then resume
Kill the process mid-conversation and start it again. The agent picks up exactly where it left off — no resume flag, no session store to operate.
Rewind and branch
Roll back to any earlier state and open the agent there. History is branchable, because history is git.
Swap models mid-task
Multi-provider with failover and per-task selection. Every provider is plain HTTP — no vendor SDKs underneath.
Gate every tool call
One policy engine authorizes tool calls and API calls alike. What an agent is allowed to do is one enforced policy, not something bolted on after.
Batteries included
Durable by construction
Agents are long-lived and stateful. Crashes resume, history branches, and code changes don't kill running agents.
Identity and authz built in
Every agent is a first-class principal. No Okta wiring, no hand-rolled RBAC.
Any model, no lock-in
Multi-provider with failover and per-task selection. Open source, so you can inspect every line.
An honest comparison
Just want to run one model's agent in your code? Use their SDK.
If your goal is simply running a single vendor's model as an agent with good tool support, the vendor's own SDK is the simpler choice — genuinely. This SDK is what you reach for when you need to trust, govern, and account for a fleet: durable agents, cross-vendor freedom, identity, and audit. Moving over is a library swap plus light restructuring, and we've written the migration path down.
Your work shows up where your buyer looks
Everything built on the SDK is instrumented natively — adoption, spend, and audit reporting arrive without another integration.
Forensics & Insights
The dashboard your leadership will ask for, already fed by the agents you built.
Common questions
It's open source (LGPL) and model-agnostic, and your agent state lives in git repositories you own. Leaving is as observable as staying.
Teams usually arrive with a PRD that reads like this SDK's feature list — durable state, per-agent guardrails, versioning and rollback, bulk model management. Read the docs first; if you still want to build, you'll build better knowing what's here.
Multiple providers over plain HTTP with failover and per-task selection. If a provider speaks HTTP and server-sent events, it's adaptable.
A library swap plus light restructuring — largely mechanical. There's a written migration reference, and a working session with an engineer covers the rest.
Bring your architecture. Talk to the team that built it.
A technical deep-dive with an engineer, not a sales call.