AllocateRecursive inference to train your own models and agents
One endpoint for every model in the catalog. Every task your agents finish becomes training signal for a model you own.
One API.Every model.
Point the OpenAI SDK at Allocate and call any model in the catalog by name. Switch models without a deploy, and fine-tune the open weight ones into a model you own.
curl https://api.allocate.network/v1/chat/completions \
-H "Authorization: Bearer $ALLOCATE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "moonshotai/kimi-k3",
"messages": [{ "role": "user", "content": "Summarise this ticket." }]
}'Your data.Your weights.
Every turn your agents take is written to RecursiveDB with the outcome that followed it: merged, reverted, approved, refunded.
Compaction turns those records into training pairs. A run fine-tunes the open weight base you picked, on the work your own team did.
The adapter is served on your tenant under your own model name. Export the weights whenever you want to take them elsewhere.
Every agent gets a computer.
Agents run on machines inside your own cell. Nothing shares a runtime with another customer.
Compute and data stay where you put them: the United States, Europe or South Africa.
The harness is a pinned build. Move your tenant forward or hold it back without touching your code.
Every tool, one contract.
MCP servers and OpenAPI specs are both inputs. Allocate turns them into a single tool contract, so an agent sees the same shape whatever it is calling and you connect a new service without shipping code.
Code mode goes further: the agent writes one short program and calls many tools in a single round, instead of burning a turn on every call.
connections: [
{
name: "jira",
mcp: "https://mcp.atlassian.com/v2/mcp",
env: "ATLASSIAN_API_KEY",
tools: ["searchJiraIssuesUsingJql"],
},
]Sign up, mint a key, and send your first request in a minute.
- One key for every model in the catalog
- A computer for every agent you run
- Experiences captured while you work
- Metered per second, no seat licence
Your own cell, in your own region, with the paperwork your risk team asks for.
- A single tenant cell in your region
- Weights and data you can export
- Security review and a signed DPA
- A direct line, not a ticket queue