Coding agents
In your terminal
- DeepSeek Harness
- Claude Code
- Codex
- opencode
- pi
- omp
Run DeepSeek Harness, Claude Code, Codex, opencode, pi or omp in your terminal — or bring Inferam into VS Code, Cursor and Zed. Nothing to reconfigure, and your code is never logged.
$ inferam claude
› Add rate limiting to the upload route and cover it with a test
● Read src/routes/upload.ts
● Edit src/routes/upload.ts +18 −2
● Write test/upload.rate-limit.test.ts
● Bash npm test -- upload ✓ 14 passed01 / Works where you code
Use Inferam from your terminal or from the editor you already have open. Same model, same account, same privacy.
Coding agents
Editors
Install once, sign in once. Claude Code, Codex and DeepSeek Harness keep their config files; opencode, pi, omp and editor setup add only Inferam's entries, and inferam setup <name> --undo removes them. Install the CLI
02 / Built for agentic coding
Native protocols, reasoning controls and fair capacity — without touching your agent's configuration.
One command and your agent runs on Inferam. Claude Code and Codex are back to normal when you quit; opencode, pi and omp keep Inferam as a provider next to your others.
Prompts, code and completions are never written to our databases or logs. A prompt cache on our own servers keeps recent prompt state for up to 7 days to speed up your agent, then deletes it.
Anthropic Messages, OpenAI Responses and Chat Completions, with streaming and tool calls, so each agent speaks its own protocol.
Pick an effort from low to max, set a thinking budget, or turn thinking off. Your agent's own controls just work.
A fair-share queue gives the next free slot to whoever has the fewest requests running. Nobody gets starved.
The model runs on GPUs we operate. No third-party inference provider ever sees your code.
03 / Privacy
Coding agents send whole files, diffs and secrets they find along the way. So we built Inferam to log none of it: no request logs, no traces, no training. The model servers' prompt cache deletes it within 7 days. We record only the numbers you need to see your own usage.
Read the privacy policyNever logged
Metadata we keep
04 / The model
A fast, open-weights model with strong tool use, served on hardware we run and tune for long agentic coding sessions.
Inferam is free during the private beta. We approve invites by hand to keep the service fast. We'll give plenty of notice before any paid plans.
DeepSeek V4.1 Flash, an open-weights model, served by us on our own hardware. Your agent's model settings are pointed at it automatically.
Whichever you already know: Claude Code, Codex, opencode, pi, omp and DeepSeek Harness all work. If you're choosing fresh, DeepSeek Harness (inferam dsh), DeepSeek's own harness, is the recommended one for the model Inferam serves. Inferam turns off its session-log upload and telemetry, so your sessions stay between you and us.
Yes. Inferam works in VS Code, Cursor and Zed, both in the editor's own chat and with coding agents inside the editor. One command sets it up: inferam setup vscode (or cursor, zed). Editor setup
Only to add itself where you'd want it. For Claude Code, Codex and DeepSeek Harness, the CLI uses environment variables and command-line flags for that one session, so running them directly behaves exactly as before. In opencode, pi and omp, it adds Inferam as a provider next to your others, so you can switch to it any time, even in plain opencode, pi or omp; nothing else changes. Editor setup likewise adds only Inferam's own entries. inferam setup <name> --undo removes them.
Yes. Each agent saves its sessions where it always does, and when the agent exits, inferam prints the command that resumes it. It's the agent's usual resume command with inferam in front: inferam claude --resume <id>, inferam codex resume <id>, inferam opencode --session <id>, inferam pi --session <id>, inferam omp --resume <id>, or inferam dsh headless --session-id <id>.
Only metadata: token counts, timings, the model, the status code and which key was used. It's kept for six months to show you usage. Prompts, code and completions aren't stored in our databases or logs; our model servers cache recent prompt state on their own disks for up to 7 days, then delete it automatically. We never train on your data.
macOS, Linux and Windows, on both ARM and x86-64. On Windows, install from PowerShell with irm https://inferam.ai/install.ps1 | iex.
Yes. api.inferam.ai accepts Anthropic Messages, OpenAI Responses and OpenAI Chat Completions requests with an Inferam API key, so most SDKs work by changing the base URL.
Requests wait in a fair queue for up to 30 seconds. If there's still no room, you get a standard overloaded error with Retry-After, which coding agents retry on their own.
06 / Private beta
We approve every invite by hand and email you as soon as you're in. Free while the beta lasts.