November 2026
- Caedos 0.5.2. The Caedos skill for Claude. Each release now carries
caedos-skill.zip, a skill you add to Claude so that when you ask for work to keep happening (watch a page, tally something weekly, send a digest), Claude sets it up in Caedos rather than as a scheduled task of its own. On claude.ai, turn on Settings › Capabilities › Create and edit files, download the ZIP from the release, then upload it in Customize › Skills. Its hash is in the signedSHA256SUMSwith the rest of the release. To update it, delete the old skill and upload the new one: claude.ai has no way to replace a skill in place. - A filter by date keeps what it should.
filter'sgt,lt,gteandlteturned both sides into numbers, so dates such as2026-11-16T23:15:58Zcompared as nothing, and a filter keeping what came after a date kept nothing, without an error. Now two numbers, or numbers written as text, still compare as numbers, and any other two pieces of text compare as text, so dates written that way order correctly. - For Caedos Cloud, which isn't open yet: a backup no longer needs room for the whole database in memory. Each hourly backup was read into memory whole as it was sent, so sending it took about twice the database's size, and once a workspace's database outgrew what its machine could hold twice over, every start sent the same backup again and was stopped again. Now a backup is streamed from disk, and sending one takes about the same memory at any size. If a run is ever stopped while sending a backup, the next start sets that backup aside and keeps your synths running, and the next hour's backup is sent as usual.
- Also for Caedos Cloud: Share says a workspace has no share link yet, and so does what your AI hears about a synth's page, since Caedos Cloud doesn't serve share pages yet. What a synth publishes is still kept. And the handbook Caedos gives AI assistants says that an API which limits keyless calls by address will now and then turn one down on Caedos Cloud, where other services share the address: that's weather, and a synth that calls such an API often should get a key of its own. On your own machine nothing changes.
- A correction to 0.5.1's notes: a synth sends through one of your channels by the channel's id (for example
email_mefor Email me), not by its name, as the handbook says. It leaves the database as 0.4.0 laid it out (schema 5), so going back to 0.5.1 opens the same database. - Caedos 0.5.1. A synth whose reasoning gives no answer now waits, and misses nothing. In 0.5.0, when a synth's reasoning gave no answer, only the actions that used what the model said were skipped. The rest ran, so a synth that remembers what it last saw, as the Page watcher blueprint does, remembered the change, and once the model came back it saw nothing new and never said what had changed. Now, when a synth's reasoning was due and gave no answer (the model unreachable, a key refused, a plan's model allowance used up, or the day's budget spent), none of its actions run on that tick and nothing is kept. Its trace says why, and its next tick sees the same change and reasons about it once the model answers. Synths that don't use a model run as before.
- When no synth can reason (your key refused or out of credits, or a plan's model allowance used up), Caedos sends one notice to your alert channel, as it does when a connection breaks. It's sent once each time reasoning stops, not at every check.
- When your AI changes a running synth so that it sends through one of your channels, the change is now kept as a new synth is: it sends through the channel's own call, with the channel's address. Before, the change was kept as written, and the synth's first send failed. Going back to an earlier version does the same.
- The handbook Caedos gives AI assistants, and its list of what's connected, now say how a synth sends through one of your channels by its id, and that a channel you haven't tested yet can be used. Before, an assistant wrote its own webhook and asked you for an address your channel already had. They also say that something you want to keep happening belongs in a synth, not in one of the assistant's own scheduled tasks.
- A send you allowed is recorded as yours: its trace says it was your Allow, from the action's own page or from the control room, and which tick held it. It used to say "unknown". Your AI, reading a synth's traces over MCP, now also sees why its reasoning gave no answer.
- It leaves the database as 0.4.0 laid it out (schema 5), so it upgrades nothing from 0.4.0 through 0.5.0, and they still open it.
- Caedos 0.5.0. A synth whose model call fails no longer acts on nothing. When a synth's reasoning failed (the model unreachable, a key refused, a plan's model allowance used up) or its day's budget for the model was spent, its actions still ran, and one that sends what the model said sent the word "undefined". Now each action that uses what the model said is skipped on that tick, and its trace says why. (0.5.1 goes further: the whole tick waits.)
- An action held for your OK now expires after 30 days, and is never sent. Until then it waited forever and kept its tick's trace with it, so a synth whose sends nobody decided made the database grow without end. A notice in the control room says how many expired and from which synths, and the tick's trace goes like any other older than 30 days. On Caedos Cloud it's the plan's number of days.
- Each action waiting for your OK has its own page, at
/approve/<synth>/<action>in the control room. It shows exactly what the action would send, with Allow and Reject. Allow sends it, and from then on that synth sends without asking. Reject drops it, and the synth keeps asking. Opening the page changes nothing. Like the rest of the control room, no synth can reach it, and Allow isn't one of your AI's tools. - For Caedos Cloud, which isn't open yet: a workspace gets an "Email me" alert channel, made once, which becomes its alert channel if you hadn't chosen one, and stays deleted if you delete it. The "waiting for your OK" alert carries the waiting action's id, so its email can link to that action's page; it still never says what the action wants to send. The quarter-hour check that synths can reason no longer counts as reasoning. When the plan's model allowance is used up, the control room says the day it renews, and when Caedos Cloud turns a workspace's model calls away, the control room says so, rather than that your key was rejected.
- It leaves the database as 0.4.0 laid it out (schema 5), so it upgrades nothing from 0.4.0 through 0.4.2, and they still open it. Before going back to 0.4.2, dismiss any notice that actions expired: 0.4.2 has no words for it, and its terminal feed fails on it.
- Caedos 0.4.2. Two fixes, both for Caedos Cloud, which isn't open yet. Nothing changes on your own machine.
- An AI that connects without using a tool no longer keeps a Caedos Cloud machine up. The machine sleeps between ticks, and once someone has used it, it stays up 5 minutes so their next step doesn't wait for it to start. Until now every request your AI sent through the connector started those 5 minutes, and some AI apps connect as each of their sessions opens, or again during a chat, whether or not they then use Caedos. Now only your AI's tool calls do, from when each call is taken up. The rest of a connection (its handshake, listing the tools, a ping) keeps the machine up only while it's being answered, so a connection with no tool call wakes it for a few seconds. You in the control room still keep it up 5 minutes, and a webhook or a share page still keeps it up only while it's being answered.
- The control room goes to your account page when Caedos Cloud stops letting it in. When your sign-in had ended (30 days after it, or a sign-out in another tab) or your workspace wasn't active, Caedos Cloud refused the control room's requests, and the control room showed each refusal as an error until you reloaded the page. Now, when a request is refused, the control room asks whether it was Caedos Cloud or your workspace that refused it. If it was Caedos Cloud, the page goes to your account page, which says why, and where you sign in again if your sign-in ended. A refusal from your workspace itself, such as a plan's limit, shows as before.
- It leaves the database as 0.4.0 laid it out (schema 5), so it upgrades nothing from 0.4.0 or 0.4.1, and both still open it. From 0.3.4 or older it upgrades as 0.4.0 does.
October 2026
- Caedos 0.4.1. Stopping Caedos no longer waits on a model that doesn't answer. Since 0.4.0 a stop waits for any of Caedos's jobs that's running, and one of them is the model key check: one tiny call to your model, to see that the key still works. That call had a model call's usual limits, 2 minutes a try and up to 6 tries, so when the model's service took the call and never answered, a stop (Ctrl+C, the service stopping, or a hosted machine going to sleep) could wait about 12 minutes, or until the host ended Caedos first. Now the check gets 5 seconds and isn't retried: the next quarter hour's check is its retry. A stop ends a check that's running at once.
- A check that runs out of time, or that a stop ends, says nothing about the key, so the banner and the health check keep what the last check found. Settings → Test key, saving a key and switching provider make the same check, so they get the same 5 seconds, and a busy model (a 429 or 529 answer) shows at once rather than after up to 15 seconds of retries.
- The hour's snapshot now comes after the ticks. Caedos holds still while it copies its database, about 1.6 seconds for our test workspace on a small cloud machine and longer as the database grows, and it took the hour's copy as it started, before it was ready. Now a copy that's due waits until no tick is running and none is due within a minute, for 5 minutes at most. On your own machine the first snapshot comes a second after Caedos starts, or once its ticks are done. A stop lets a copy that's still waiting go, and the next start takes it. Each clock hour still gets one snapshot.
- In hosted mode, for Caedos Cloud, which isn't open yet, a machine woken for a tick now starts, ticks on time, then takes the hour's snapshot, sends it and sleeps; before, the first wake after each hour took the copy before it was ready. A sleep waits for the copy, and each snapshot is still sent once.
- It leaves the database as 0.4.0 laid it out (schema 5), so it upgrades nothing from 0.4.0, and 0.4.0 still opens it. From 0.3.4 or older it upgrades as 0.4.0 does.
- Caedos 0.4.0. A restart no longer pushes every synth back a full interval. A scheduled synth's next tick now comes from its last one, so an hourly synth restarted at :40 still ticks at the hour it would have, and a daily one keeps its time of day across restarts and upgrades; before, a daily synth could skip a day. A synth whose tick came due while Caedos was stopped ticks once as soon as Caedos starts, however many it missed. Run now, webhooks and signals don't count as its last tick, and a synth that has never ticked on its schedule still waits an interval from the start.
- A synth patched or reverted while one of its ticks was running no longer ticks twice each interval from then on.
- Caedos's own jobs (the self-healing check, clearing old traces, the hourly copy of the database, the digest's check, the model key check and the connection health checks) now each run once in each period of the clock: each 5 minutes, quarter hour or hour, in UTC. The database keeps when each last ran, so a restart doesn't run them all again: the key check no longer calls the model at every start, and the self-healing check no longer waits 5 minutes after one. Stopping Caedos waits for a job that's running. An hour with nothing else going on now gets its hourly copy too; Caedos still keeps 48 hourly copies and 7 daily ones.
- A repair Caedos is watching carries over a restart: the change it's trying stays, and is judged after the ticks it watches. Before, a restart undid it. The limit of three tries for one problem holds across restarts too; before, a restart started the count again.
- The control room catches up at once when you come back to its tab, rather than at its next update a few seconds later. While its tab is hidden it sends nothing, as before.
- In hosted mode, for Caedos Cloud, which isn't open yet, a machine now stops itself between ticks and Caedos Cloud starts it again for the next one. Ticks fall on the workspace's grid, so a workspace wakes at most once per step of it, 288 times a day at 5 minutes, whatever its synths. Once nothing is due in the next minute, nothing is running and no one has used it in 5 minutes, it tells Caedos Cloud when to wake it, at its earliest tick or a day on, and stops once that's answered. If the call fails, it stays up and asks again. That needs two new settings,
CAEDOS_WAKE_URLandCAEDOS_WAKE_TOKEN, without which a hosted machine doesn't start. - Hosted, only you in the control room and your AI through the connector keep the machine up for 5 minutes; a webhook or a share page keeps it up only while it's answered. Caedos's jobs count their periods from the grid, so a wake runs each one that's due once. Each snapshot is sent once, not again at every start, and a send gives up after 30 seconds plus a second for every 100 KB and is tried again at the next wake. The control room stops updating after 5 minutes with no mouse, key, scroll, touch or focus in it, says "Paused while you're away", and catches up when you're back.
- The first time 0.4.0 opens your workspace, it saves a copy of the database in the
snapshotsfolder beside it, then upgrades it from database schema 3 to 5 (when each of Caedos's jobs last ran, and the repairs in progress). 0.3.4 and older can't open the upgraded database. To go back, stop Caedos and runcaedos restore <the copy's name>with the older version (caedos restorealone lists the copies): the workspace comes back as it was before the upgrade, with every synth paused. - Caedos 0.3.4. When a name lookup fails with "try again" (
getaddrinfo EAI_AGAIN: DNS couldn't answer just then), Caedos now treats it as a passing outage, like a timeout. A synth rides out 8 failed ticks of it in a row before Caedos stops it, not 4. - When a source fails on a tick, what it would have said is unknown, not changed. A check that reads it doesn't run, so a failed fetch no longer sets off an "it changed" alert, and nothing computed from that source runs that tick. The tick is recorded as idle, with the source's error. A value the synth keeps to compare with is kept, so the next good tick neither alerts again nor misses a real change. A synth meant to act on a failure can still read
$sources.<id>.error, and a synth set to reason on every tick still does. - When a synth stalls or its source changes shape, Caedos now asks for a rewrite at most twice for one problem, whatever comes back: a rewrite it used, one its rehearsal refused, one it couldn't check, one it had already tried, or nothing usable. Before, an answer with nothing usable stopped the synth at once, and a rewrite Caedos couldn't check was asked for again every five minutes with no limit. Now an empty answer is asked for again at the next five-minute check, and each try is its own line in the Live Feed.
- When Caedos stops a synth it couldn't repair, the autopsy reads only what happened since Caedos last stopped it, so it no longer quotes old failures or counts an earlier problem's repairs. It says what Caedos tried in plain words, such as "rewrote it from what I saw — the rewrite failed its rehearsal, so I didn't use it". The daily digest's repair lines are in plain words too.
- It leaves the database as 0.1.1 laid it out, so it upgrades nothing from 0.1.1 through 0.3.3, and they still open it; 0.3.3's Live Feed shows only the synth's name for the three new kinds of repair line. As with 0.3.0, a version before 0.3.0 can't send a call held for your OK. From 0.1.0 it upgrades as 0.1.1 does.
- Caedos 0.3.3. A synth that waits for something, such as a number crossing a line, text matching a pattern or a new item in a list, is no longer stopped as stalled while the value it watches moves and it rightly stays quiet. Only a synth that fires on any change in what it watches (
fast_path_diff) is called stalled when that value moves and it doesn't fire. For every synth, a watched field that comes back empty while its source still answers, and reasoning that keeps failing, still count as a stall. - When a synth has stalled for three ticks in a row and has no change threshold Caedos can loosen, Caedos now rewrites it from what it saw: which watched value moved or came back empty, on how many ticks, and what the source returned. Before, it stopped the synth without trying. Each rewrite is one call to your premium model tier (Claude Sonnet 4.6 by default), or up to five when an answer fails Caedos's checks (ten for a synth on the Relay whose rewrite drops its topics). Caedos rehearses the rewrite before using it, which calls a model only if the rewritten synth reasons, and drops one the rehearsal finds broken. A rewrite never loosens the synth's leash, and if the next ticks don't recover, Caedos goes back to the version before. After two rewrites for one stall, or when no usable rewrite comes back, Caedos stops the synth as before and says what it saw.
- In hosted mode, for Caedos Cloud, the clean-up of old traces logs how many traces, decided approvals and signals it deleted and how many traces it trimmed, and a part that fails is logged as an error.
- It leaves the database as 0.1.1 laid it out, so it upgrades nothing from 0.1.1 through 0.3.2, and they still open it. As with 0.3.0, a version before 0.3.0 can't send a call held for your OK. From 0.1.0 it upgrades as 0.1.1 does.
- Caedos 0.3.2. The five-minute check on your synths no longer asks a model whether each one's recent work still fits its goal. That answer never started a repair, and on our test workspace it cost ten times what the synths' own reasoning did. The checks that catch a stalled or broken synth use no model and still run, so the check now calls a model only to repair a synth it found broken, or for drift detection if you turned it on in Settings.
- Past a synth's newest 20 traces, a trace whose tick didn't act keeps only what its fast path read from each source, with the answer's size and a fingerprint, instead of the whole answer. The newest 20, and traces of ticks that acted or that a call waiting for your OK came from, stay whole. Caedos trims the older ones as it starts, then every hour, so the database grows more slowly. Replay on a trimmed tick says it sees only what the fast path read.
- Caedos now runs on Bun 1.4.3 (it was 1.3.14).
- It leaves the database as 0.1.1 laid it out, so it upgrades nothing from 0.1.1 through 0.3.1, and they still open it; traces it already trimmed stay trimmed. As with 0.3.0, a version before 0.3.0 can't send a call held for your OK. From 0.1.0 it upgrades as 0.1.1 does.
- Caedos 0.3.1. In hosted mode, for Caedos Cloud, sending a backup over HTTPS no longer crashes Caedos. Nothing changes when Caedos runs on your own machine.
- Caedos 0.3.0. Caedos now has a hosted mode, for Caedos Cloud, which isn't open yet. It's off unless
CAEDOS_HOSTED=1is set; with it off, nothing about how Caedos runs on your machine changes. - A secret a service echoes back, such as "invalid token: Bearer …", is now masked as
[secret NAME]in everything Caedos records: traces, state, signals and calls waiting for your OK, so in the control room and in snapshots too. A synth that wrote one of your secrets into its own state now reads back the mask. - When you Allow a call, Caedos fills in only the
$secretsrefs its author wrote. Text that a source, a model, state or an event put in the call is sent as it was held, even when it looks like a ref. A call already waiting for your OK when you upgrade is filled in the old way. - Stopping Caedos waits for requests still running, such as Run now or an Allow being sent, so they finish before the database closes. While it waits, new requests get "Caedos is stopping".
- It leaves the database as 0.1.1 laid it out, so it upgrades nothing from 0.1.1, 0.1.2, 0.2.0 or 0.2.1, and they still open it. An older version can't send a call that 0.3.0 held for your OK, so decide those before going back. From 0.1.0 it upgrades as 0.1.1 does.
- Caedos 0.2.1. On Windows, a console window no longer opens each time a synth uses an MCP server that Caedos runs itself, one with a command such as
uvx mcp-telegram-botornpx -y @modelcontextprotocol/server-github. When a call is done, Caedos now ends that server and anything it started; before, one could keep running, its window open, long after the call. - An MCP server that doesn't answer within 30 seconds now fails the call, and Caedos ends it; before, the call could wait for it forever.
- It leaves the database as 0.1.1 laid it out, so it upgrades nothing from 0.1.1, 0.1.2 or 0.2.0, and they still open it. From 0.1.0 it upgrades as 0.1.1 does.
- Caedos 0.2.0. Its MCP tools have new names. The ten you use to write and run synths are now
synth_validate,synth_shadow,synth_deploy,synth_run,synth_ps,synth_get,synth_traces,synth_patch,synth_pauseandsynth_kill;caedos_devices,caedos_blueprintsandcaedos_docskeep theirs. The old names aren't kept as aliases, so a permission rule your AI client saved needs the new name: in Claude Code,mcp__caedos__caedos_runbecomesmcp__caedos__synth_run. - Each tool has a title and says whether it only reads, changes or removes something, so a client can ask before the ones that change things.
/mcpspeaks MCP 2025-11-25, 2025-06-18 and 2025-03-26 as well as 2024-11-05, and answers each client in the version it asks for. - It leaves the database as 0.1.1 laid it out, so it upgrades nothing from 0.1.1 or 0.1.2, and they still open it. From 0.1.0 it upgrades as 0.1.1 does. Traces from before keep the tool name they were recorded with.
- Caedos 0.1.2. It leaves the database as 0.1.1 laid it out, so it upgrades nothing and 0.1.1 still opens it. From 0.1.0 it upgrades as 0.1.1 does.
- The hourly sweep that deletes traces older than 30 days now deletes decided approvals and Relay signals by the same rule: each synth keeps its newest 100 allowed or rejected actions, each topic its newest 100 signals, and an action still waiting for your OK is never deleted.
- A synth that reads a Relay topic with
signal_subscribenow asks the database for that topic alone, so a busy topic slows it far less. - Caedos 0.1.1. The first time it opens your workspace, it upgrades the database and keeps the copy from before until you delete it. 0.1.0 can't open the upgraded database; the release notes say how to go back.
- Traces keep each source's observation under about 64 KB, older ones too, so a big workspace's database shrinks once on the upgrade. Snapshots keep the newest 48 and one a day for a week.
- A running Caedos holds its workspace: a second
starton it refuses, andmcp,watch,ps,export,importanddoctorfind it whatever its port. - A restore shows a notice in the control room. The Today line counts what was sent, the daily digest says what each synth sent, and Settings says synth where it said process.
- Caedos 0.1.0, the first release. One program for Windows, macOS and Linux, with nothing else to install.
install.shandinstall.ps1fetch it and check it against the release's checksums, which every release signs. - The
caedoscommand:start,mcp,service install,watch,ps,export,import,restore,doctor.caedos mcpconnects any MCP client, over stdio, to the Caedos that's running. - One data folder per computer, hourly snapshots, and a backup before every database upgrade. Traces older than 30 days are deleted, keeping each process's newest 100.
- Sources can't write without Allow, no synth can reach Caedos's own API, and pages from other websites can't drive it.
- A notice when the computer slept, and one line to your alert channel when an action waits for your OK.
- Aether is now Caedos, said KAY-dos. The MCP tools are
caedos_*and the variablesCAEDOS_*. - The license is now the Elastic License 2.0: free for personal use and for work. Caedos is closed source, and its releases are binaries.
September 2026
- Everything an operator reads now calls a running synth a process.
- The API listens on localhost unless
CAEDOS_HOSTsays otherwise, and warns when it doesn't. - Traces record provenance, what caused a tick and what it was allowed to do, instead of guessing it.
August 2026
- Caedos became a kernel with syscalls. The MCP server became the way your AI installs and changes processes (
validate → shadow → deploy → run → traces). - The built-in conversational builder was removed. Your assistant is the author now.
- The control room's home is the process table; the Live Feed is the day's look; the process row opens traces.
- The first tick of a change detector is observe-and-remember.
- Deterministic arithmetic: numbers are computed, never left to a language model.
- Self-repair hardened: interrupted repairs never adopt an unproven candidate; failures are split into weather and defects.
- Workspace policy moved into Settings.
July 2026
- REST connections as a first-class protocol.
formatandsorttransforms: lists become text without a model.- Build journal and debug bundle for when a build won't come out right.
- Model-key health: surfaced, probed and recovered.
June 2026
- The declarative runtime: synths, the six-phase tick, primitives, stores.
- Cost and budget governance, change history, workspace export and import, alert channels.
- Unified Connections for MCP servers, with drift alerts.
rss_readprimitive. Shadow rehearsals with real reasoning.
May 2026
- Project started.