Temperis is the AI-native code editor that gives every model — local or cloud — senses, memory, hands, and a conscience. Grounded context. Verified verification. Compounding memory. The app is free.
Free — no trial clock, no card · 141 agent tools · works fully offline with local models · lifetime license
your agent debugs your code live
Model quality sets the ceiling. Temperis's job is to make every model — a 3B local model or a frontier cloud model — operate at its ceiling.
A hybrid semantic + keyword code index, live type information, references, a project graph, runtime traces, and screenshots of the app it's building — fused into one queryable brain.
Lessons, dead-ends, procedures, taste, and a run-outcome ledger — captured automatically, distilled into a task primer at turn zero. Your editor gets measurably better at your codebase.
141 tools: files, git, terminals, web, live app hosting, OS-level desktop control, sub-agents, and tools the agent writes for itself mid-task.
A finish gate that demands proof, an adversarial skeptic that hunts counterexamples, risk-calibrated permissions, and checkpoints under everything. "Done" means proven done.
Every capability below is a real subsystem with its own service, tools, and settings. They apply to every provider identically.
When the agent says "done", Temperis forges a dozen broken versions of its own fix and checks whether your tests can even tell them apart. A test that can't catch a planted bug is a hollow test — and Temperis refuses to finish on one. The first editor where "verified" is itself verified.
the Crucible · ships free with the appThe intelligence that keeps Temperis improving runs on our servers, not your disk, so it gets smarter without a reinstall. The app is the body — your files, terminals, tests, local models; the Brain is the mind. Your code is processed in transit and never retained. Optional subscription.
Temperis Brain · optional subscriptionEvery file write instantly pulls live diagnostics plus the runtime errors the running app actually hit. A run that edited code cannot finish until the finish gate re-verifies every touched file, runs the affected tests, and gets a one-line answer to "how did you verify this against the goal?" Failures bounce the model back with the exact errors.
Speculative edits that never touch your tree until proven. Edits stage in memory, the TypeScript server verifies all staged files together, then commit lands everything atomically under a checkpoint — and auto-rolls back if tests fail. Failing changes simply never persist.
One tool call — ask_codebase — answers "how does X work / what calls Y" by fusing index search with live LSP types, reference counts, and graph edges into a cited, token-minimal answer.
ask_codebase · GROUND TRUTHBefore a substantive edit may finish, a second model is handed the diff and told to find one concrete input that breaks it. A real counterexample blocks completion.
adversarial verify · on by default"Keep working on X until done" — persisted to disk, surviving app restarts. On relaunch Temperis offers to resume, the agent picks up its own plan, updates progress, and closes the goal only when it's truly met.
campaign_start · auto-resumeLessons, dead-ends ("AVOID — already failed here"), procedures, and taste — auto-captured, auto-recalled as a ~300-token primer at turn zero.
Every run logs outcome, duration, gate bounces, and backtracks. The agent can query its own track record — and gets warned when similar goals recently failed.
The agent drives the app it just built — click, type, wait, assert — inside the IDE, and gets a per-step pass/fail transcript. Proof the button works, not just renders.
Trace-instrumented runs record real arguments, returns, exceptions, and per-line locals. The agent debugs from recorded reality, not speculation.
Stagnation and repeated gate failures trigger an automatic checkpoint restore and a strategy change — bounded, never destructive.
Risk-scored permissions: evidence-backed low-risk edits flow; risky or irreversible actions ask. Explicit deny rules always win.
The agent writes its own sandboxed tools mid-task and can call them the same run — and every run after.
Proven tool sequences replay deterministically with zero model calls. The model is only re-engaged if a step fails.
Not logs. Not guesses. The agent sets a conditional breakpoint, freezes the real process, reads every live variable — then changes one to test its theory, and lets execution keep going. Watch:
That's debug_session_start → scopes → eval → continue — four of the 141 tools every Temperis agent carries. Ships free, works with local models.
Real tool output from a Temperis session — the verification transcripts every run has to survive.
Most AI editors reason about what your code probably does. Temperis gives the agent instruments that measure what it actually does — and wake it when reality changes.
"Wake me when this file changes, when the dev server logs an error, when the type errors flip — including to zero." The agent stops burning turns on checking.
Record a run, edit, record again. The diff names exactly which function's runtime behavior moved — and stays silent about everything that didn't.
Real, trusted keyboard input with frame-level timing — hold Right 500ms, tap Space 80ms, screenshot where the player landed. Canvas games included; DOM bots can't touch this.
Anonymous, content-free start/update/crash signals from real installs — so a broken release gets caught in hours, not whenever someone emails. Off-switch included.
Every verified piece of work auto-deposits a decision ledger — what changed, why, how it was proven, what failed on the way. The next session recalls it instead of rediscovering it. Dead ends stay dead.
Memory, knowledge, dead ends, run outcomes, session history, git — one query. And the deposit is structural: the harness writes it when verification passes, whether or not anyone remembers to.
Files, code intelligence, git, terminals, a live debugger, tracing, watches, recall, verification, memory, campaigns, desktop control, MCP — everything an agent needs to actually finish the job.
Fully offline with local models, or plug in any cloud key. A model router sends mechanical work to cheap models so your frontier model only reasons.
Every editing turn snapshots touched files first. Anything can be rolled back — one click, one tool call, or automatically.
Per-category levels, fine-grained rules like deny: Edit(**/.env),
permission modes, and risk-calibrated auto-allow. Deny always wins.
A local model screens every terminal command for danger before it runs.
The agent must read a file before overwriting it; stale-content edits are refused outright.
Conversation archives redact secret-shaped content; env files are excluded from indexing; sensitive paths can be denied outright.
Agent-authored tools run in a Web Worker with a read-only API and a save-time denylist.
Security boundaries fail CLOSED — a permission check that cannot be evaluated denies. Verification layers degrade openly and say so, rather than trapping a run — and every "done" has to survive the finish gate first.
| Capability | Temperis | Subscription AI editors |
|---|---|---|
| Pricing | Free — the app is yours, full stop | $20+/mo, forever |
| Local models / fully offline | Yes — first-class | Limited or none |
| Mandatory verification before "done" | Finish gate + tests + runtime truth | Model says "done" |
| Adversarial self-review | Built-in skeptic | — |
| Speculative edits with auto-rollback | Shadow transactions | — |
| Goals that survive restarts | Durable campaigns | — |
| Memory that compounds per-project | Lessons, dead-ends, taste, outcomes | Partial |
| Drives & tests the app it builds | Live Apps + interactive verify | Screenshots at best |
| Voice + OS desktop control | Jarvis | — |
| Roblox bridge | Built-in | — |
| Your data leaves your machine | Only if you choose cloud models | Always |
Temperis is a free download — no license key, no trial clock, no card. The Temperis Brain is the only paid product: an optional subscription that adds server-side intelligence. Bring your own API keys — or pay $0/token with local models.
A $20/month editor costs $720 over three years — and the editor itself was never the hard part. Temperis gives you the whole editor free; the Brain is one optional $39/mo subscription you can cancel anytime.
By downloading, starting a trial, or purchasing, you agree to the Terms of Service, the EULA, and the Privacy Policy.
Yes. Point Temperis at your local Ollama and everything — the agent, the code index, embeddings, the command safety judge, voice transcription — runs on your machine. Cloud providers are optional, added only if you paste a key.
Really free — download, install, use, forever. No license key, no trial clock, no account, no card, and all future app updates included. The only paid product is the optional Temperis Brain subscription ($39/mo, cancel anytime), which adds server-side intelligence on top of the free app. If you bought Temperis back when it was $99: thank you — your license and everything it covered stay honored, and you now also get everything free users get.
No — the free app is complete on its own and runs fully offline. The Temperis Brain is an optional subscription add-on ($39/mo) that runs the always-improving intelligence on our servers, so the app keeps getting smarter without a reinstall. The app is the body; the Brain is the mind. Your code is processed in transit and never retained. Cancel anytime.
Any Ollama model, Anthropic's Claude family (with extended thinking and vision), anything on OpenRouter, and your installed Claude Code CLI as the agent. A router can split work across tiers — frontier model for reasoning, cheap model for mechanical edits and synthesis.
Everything Temperis learns — memory, lessons, outcomes, checkpoints, campaigns — lives in
a .apex/ folder inside your workspace, on your disk. Code only leaves
your machine if you choose a cloud model, and conversation archives redact
secret-shaped content.
It has to get through seven layers first: checkpoints under every edit, layered permissions with deny rules, a command safety judge, read-before-write enforcement, the finish gate, the skeptic, and — if you enable shadow transactions — edits that never touch disk until proven green.
Windows 10/11. For local models: Ollama plus enough VRAM for your chosen model (Temperis checks and warns before loading anything too big). Cloud models work on any machine.
7 days, no questions asked. Email support@temperis.pro.
Free download · Windows · offline-capable · Brain optional at $39/mo