Versioned app config
Push managed snapshots with agents, task types, skills, capabilities, defaults, graders, and pipelines.
Managed agent backend
Ship AI agents like product infrastructure.
Trellis gives your app a control plane for agent execution: versioned config, runtime choice, automations, evals, pipelines, mesh coordination, and the telemetry to operate it all after launch.
One backend
runs, chat sessions, schedules, webhooks, evals, and CI pipelines
Runtime flexible
Claude, OpenAI, Codex, DeepAgents, Z.AI, DeepSeek, CCR, and mock
Operational by default
logs, costs, traces, approvals, artifacts, and health checks
Platform
A control plane for every agent your product depends on.
Apps sync snapshots into Trellis. Trellis resolves the right prompt, runtime, tools, budgets, context, and task backend at execution time.
Dynamic agent runtime
One execution path resolves per-agent models, runtime overrides, capability tags, task templates, and resume state.
Tooling without rewrites
Expose backend routes as capabilities, opt agents into built-in tags, and add app data without changing Python code.
Scoped execution
Per-context concurrency, API-key app isolation, budget limits, approval gates, and hardened webhook entry points.
Operate
Move from prompt demo to dependable workflow.
Automations that know why they fired
App events, cron schedules, and inbound webhooks all become traceable Trellis runs with stable trigger identities.
Operator console built for real traffic
Inspect tasks, live logs, run history, cost windows, prompts, knowledge, agents, app config, and health checks.
Regression loops for agent quality
Turn failures into datasets, run graded sweeps, compare model or prompt variants, and triage cases with durable notes.
Mesh and pipelines for bigger jobs
Coordinate agents through shared work, dispatch sandboxed CI jobs, stream logs, validate outputs, and emit lifecycle events.
Why teams pick Trellis
It treats AI as a system you can inspect, test, and improve.
Trellis gives every run a home: the prompt that launched it, the snapshot it used, the events it emitted, the artifacts it wrote, and the operational trail your team needs when the answer matters.
Run telemetrystatus, cost, turns, duration, traces
Chat sessionsstateful agent conversations with resume
Approvalsoperator gates for sensitive pointer actions
Artifactstyped HTML, JSON, text, PDF, image, and video payloads
Outbound eventssigned callbacks for app workflows
Admin PWAmobile-aware console with vendored assets
Quickstart
Register an app, sync a snapshot, run an agent.
Trellis stays close to the command line and API surface your engineering team already uses. Start local, then keep the same primitives in production.
# Install and migrate the control plane
just install
trellis migrate
# Start Trellis and publish app config
trellis serve
trellis app sync --config .trellis/app.yml --show-api-key
# Run, automate, evaluate, and observe
trellis run workspace chat --app my-app --prompt "Summarize today's risk"
trellis event "order.large_batch" '{"order_id":"ORD-123"}' --app my-app
trellis evals sweep runtime_parity --app trellis-self