Skip to content

noetaThe Claude Code–compatible coding agent you can self-host on many workers.

The model sees what Claude Code shows it. Underneath, every task is an append-only log with one writer in a Postgres (or SQLite) store, so any worker continues any task, a task waiting for a person holds no process, and a crash or a redeploy only moves the task to another worker.

noeta

How it fits together ​

Ways in Terminal UI noeta Headless, Agent SDK noeta -p Editors (ACP) noeta acp HTTP / WebSocket noeta serve Python noeta.Client noeta What the model sees Claude Code's tools, schemas, prompt rules, reminders, compaction, request shape Plugins: one Extension interface fs shell web permissions hooks skills subtasks mcp compaction built-in ones load exactly like yours Kernel Client create · send answer · watch Runtime claim · renew release TaskRun fold · step append Store sqlite or Postgres one event log per task an inbox per task claims with an epoch snapshots Model provider anthropic · openai_compat openai_responses append · fold request · reply

Every front end talks to one client API: create a task, send to it, answer it, cancel it, read and watch its events. Workers claim tasks from the store, fold each task's log back into its state, call the model, run tools through plugins and append what happened. What the model sees is built by the plugins to match Claude Code. How it works →

What sets it apart ​

Many workers, one store. Point any number of noeta worker processes at one Postgres store and put noeta serve in front. A worker claims a task for one step at a time and appends what happened to the task's log, so the next step can run on any worker: a deploy, a crash or a lost machine only moves the task. Front ends attach to any server with the same API (Run it as a server, Deploy with Docker).

Claude Code's contract on a server. Models are tuned against Claude Code's tools and prompts, so noeta copies what the model sees, not how Claude Code is built: tool names and schemas, the order and roles of prompt parts, reminders, compaction, the request a provider gets. make parity replays recorded Claude Code sessions against noeta offline, compares every request, and reports each difference not listed as accepted with its reason (Matching Claude Code). Your settings, hooks, skills, agents, MCP servers and plugin folders keep their Claude Code format but live in ~/.noeta and .noeta/; noeta never reads ~/.claude (Configuration).

An event log with one writer. A task is a sequence of ten event types in the store. A worker must claim a task before it writes, and every append names the sequence number it expects and the claim's epoch, so a worker that froze and lost its claim can never write again. A tool call is logged as started before it runs; after a crash the next worker answers an unfinished call "outcome unknown" instead of running it twice (A task is a log, Crash recovery).

Waiting holds no process. A permission prompt, an AskUserQuestion, a sub-agent or a timer parks the task: it logs Waiting, releases its claim and is gone from every worker until its inbox gets the answer. A waiting task is a row and a log in the store, with no thread or process behind it (Waiting and waking).

Everything is a plugin. The kernel runs logs; every capability is a plugin behind one Extension interface, found by name through Python entry points or as a single file in ~/.noeta/plugins/. A task records the plugins it started with, and a worker missing one never runs the task without it (The plugin system, Write a plugin).

Private by default. noeta sends nothing anywhere except requests to the model provider you configure: no telemetry, no log upload, no update check. Tools reach out only when the model calls them, under your permission rules.

Measured next to Claude Code ​

+5.4
SWE-rebench 07-2026: noeta − Claude Code, standard error 3.6
111 tasks, 50 min cap, two runs pooled · noeta 50.5 %, Claude Code 2.1.312 45.0 % · on par or slightly ahead, not significant
+1.9
Terminal-Bench 4.0: noeta − Claude Code, standard error 3.2
54 of 66 tasks, 2 h cap, one run · noeta 3.7 %, Claude Code 1.9 % · near zero for both on this model

noeta 1.0 and Claude Code 2.1.312 driven by the same model, DeepSeek V4.1 Flash (effort high), not a Claude model; same tasks, caps and machine. The cost per run was in the same range ($5.01–$5.77 against $5.70–$6.25 on SWE-rebench). Method, every run and the caveats →

Where to go next ​