Run your agents on FlexAI
Bring the skills and tools your agents already use, and FlexAI's harness runs them on open models, governed and fully audited.
Task
Review PR #72 and draft release notes
- Model
- Qwen / DeepSeek
- Tools
- GitHub · Docs · Linear
- Controls
- Approval required before write
- Trace
- 12 steps · audited
Everything you already built
We make it run better.
Different doors. Same runtime. No vendor lock-in
Hand FlexAI a skill or an intent through the API, the SDK, the shell, or Slack. Different doors, one runtime, the same run object back. Endpoint detail comes with the trial.
Runtime independence
Skills run on Claude Code, Codex, and Cursor without modification, and plug into custom Claude or OpenAI API loops.
Model independence
Claude, GPT, DeepSeek, Llama, Qwen, all behind one router.
Hardware independence
NVIDIA and AMD, equal-status.
We understand what you bring and make it ready to run
You decide what to run. FlexAI shows what is ready.
Instruction import
AGENTS.md, prompts, playbooks, and operating doctrine.
Skill registry
Detect capabilities, dependencies, and triggers.
Tool inspection
Verify tools, permissions, and side effects.
Secret binding
Connect tokens and accounts without putting them in model context.
Trust scan
Policy, safety, data-sensitivity, prompt-injection, and supply-chain checks.
Compatibility score
What works, what is missing, and a confidence score.
The harness for production agents
Every run is routed, scoped, tool-gated, approval-aware, and traced.
Router
Routes intent, surface, skill, policy, and the model/runtime path. Declare your constraints; the router picks per run.
Scoped access
Enforces access policy, sensitivity, and least privilege, with the denial reason shown before any content is fetched.
Sensitive retrieval
Loads the right context at the right time, sensitivity-aware and clearance-aware.
Tool gateway
Executes registered tools and connectors through governed, auditable calls.
Correction memory
Captures fixes and learns from corrections, so your agents stop repeating the same mistakes.
Audit & approvals
Append-only trace of every action, with human approval gated on sensitive or mutating steps.
Underpinned by Model Council: picks the best model and runtime for each run across quality, cost, readiness, and fallback.
Every run returns proof you can audit
Traceable, measurable, and provable.
Artifact
The final output: markdown, JSON, files, reports.
Model and lane used
Which model, lane, runtime, and fallback ran.
Context loaded
Instructions, memory, sources, and data used in the run.
Tools called
Which tools ran, and what came back.
Proposed writes
What would change and why. Proposed, never automatic.
Approval state
What was approved, rejected, or awaiting a decision.
Audit trace
An immutable trace of every step, decision, and approval.
Compatibility score
How well this agent runs on FlexAI in this policy and model lane.
What accumulates for you
Every run makes the next one better.
Your corrections persist across sessions.
Retrieval sharpens on your data.
Your audit history compounds.
Your tool topology deepens.
Agent SDK trial: now open
We're onboarding a limited set of teams running real agent workloads. Bring a skill and a workload; we'll get it running on FlexAI primitives and measure the results with you.
Limited seats. Teams with production agent traffic get priority.
Tell us about your agent workload (your company, current harness, and rough volume) and we'll be in touch.