Agent runtime — Governed model routing
ClearCode Fusion is two things working as one: a local agent runtime that executes tools on your own machines, and a gateway that routes every turn across local and hosted models under a single policy you control. Cloud is optional. It is never required.
“An agent that cannot act is a chatbot. An agent that acts without a receipt is a liability. You need both halves.”
The platform
Agents execute real tools on the host they run on: shell, file read and write, database queries, and your own internal services. Scoped to configured roots, with an approval gate on anything that mutates state.
One endpoint in front of every model you use, local and hosted. Per-seat preference order, automatic failover, circuit breakers, and a spend policy enforced at selection time.
An agent is its role, memory, and record — not the model serving the turn. Swap the engine underneath; the agent, its authority, and its history stay intact.
Agent runtime
Agents do real work on real systems. That is only safe because every capability is bounded, and every action produces a record.
File and shell access is confined to configured directories. An agent cannot wander the filesystem because it was told to.
Read-only tools execute immediately. Anything that mutates state goes through an approval gate before it runs.
An agent can only call tools that were registered for its seat. A missing capability is reported as a gap, not worked around.
Every tool call returns a receipt: what ran, what it returned, and whether the intended effect was actually proven.
Model gateway
Governed routing
Each agent seat carries an ordered model preference matched to its work. The gateway honours that order, and one policy governs every seat rather than a config file per agent.
Resilience
Circuit breakers trip on timeouts and server errors, and the turn cascades to the next model in order. A provider outage degrades one route, not the whole operation.
Spend control
Self-hosted models carry no per-token cost and sit at the bottom of every cascade. Routine work can run entirely on hardware you already own; hosted models are used where the task warrants them.
Provider choice
The gateway is provider-neutral. Add the models your organisation already has agreements with, set the order you want, and change it without redeploying your agents.
Evidence
We run our own company on this platform, so its controls were shaped by real operations rather than by a demo.
A command can exit cleanly and still accomplish nothing. Every receipt distinguishes “it ran” from “the intended effect is proven,” and an unproven effect is reported as unproven.
An agent claiming it completed something is not evidence. The receipt, the row, or the artifact is. Claims without a receipt are flagged rather than accepted.
Changes are staged and tested before they replace anything live, and every change carries a rollback path. Rollback is part of the design, not an apology.
Consequential actions stay visible and subject to approval. Money, publication, and destructive operations remain human-gated by default.
Runs inside your network, on your hardware, under your rules. Your data and your models stay yours.
Get started
Tell us what your agents need to do and what they must not touch. We will map the runtime boundaries and the model policy, and show you the receipts.
Talk to us about ClearCode Fusion