Repository navigation
AgentSwarms 1.7.0 is the first release published as images, and the first that deploys
to AWS without assembling it by hand. Every image is built for amd64 and arm64,
smoke-tested, scanned and signed. An install picks a release with AGENTSWARMS_VERSION,
and rolls back the same way. The schema now installs on a plain PostgreSQL such as
Amazon RDS. deploy/aws deploys the whole platform with Terraform: as VMs behind a load
balancer, on Amazon ECS, or on EKS from the new Helm chart. Users' code runs on Fargate
there, or on a node pool of its own. Behind a load balancer, restarts and rolling
deploys no longer cut replies off, and the scheduler runs once across the fleet. The
lakehouse gains Iceberg v3, time travel and Iceberg targets on the Spark engine, and the
catalog reads Hive Metastore and AWS Glue. On security: credentials, traces and
approvals are sealed under per-user and customer keys, sandboxes are isolated from one
another, and a dynamic scan of the running app found nothing exploitable. Read Upgrading
before you move: several migrations must be applied in the order given there.
Everything 1.7.0 changes, adds and fixes is in CHANGELOG.md: the whole of it is longer than a GitHub release may be.
Upgrading
- Apply
20261114000000before choosing the ECS sandbox backend (R477). It lets
the runtime settings nameecsand adds the three columns and the function a Fargate
sandbox fetches its environment through. It only adds things: either order with the
build is safe, and the Docker and Kubernetes backends never touch them. - The images have new names (R470):
ghcr.io/agentswarms-fyi/<image>:localfor a
build from the checkout, where they wereagentswarms/<image>:latest. Re-running
scripts/setup.sh(ordocker compose up -d --build) builds them under the new
names. The installer clearsNOTEBOOK_RUNTIME_IMAGEwhen.envstill holds the old
default,agentswarms/notebook-runtime:latest; do it by hand otherwise, or the app
keeps starting kernels from the old name. The old images stay on disk until
docker image prune.setup-k8s.shdeploys the new names; a cluster that loaded
the old ones needs the new ones loaded the same way. - Update the notebook gateway image with the app (R466). Sandboxes the new app
starts refuse a gateway that sends no token, so an old gateway in front of a new app
cannot open notebook kernels. The new gateway works with sandboxes an older app
started, anddocker compose up -d --buildupdates both. Running MCP servers and
scorers keep working until they restart; the ones started after the update check the
token. - Apply
20261111000000first, then deploy the build (R456). The catalog page now
reads the newmetadatacolumn by name, so the new build fails on a database that does
not have it yet. The migration only adds things, so the running build is unaffected by
it. A bucket with Hive-partitioned folders is re-grouped on its next crawl: the
per-folder assets are replaced by one table, and anything written on those old assets
by hand goes with them. - Deploy the build first, then apply
20261110000000(R454). The migration drops
the two functions the browser used to seed sample datasets with. In that order nothing
breaks. The other way round, a new instance can't seed its samples until the build
arrives, and existing samples are unaffected either way. Run
npx supabase db push --remotefrom the repository folder, because the link to your
project is saved there. To confirm it
landed withoutpsql, see the schema audit.
Before you upgrade, or after, seedocs/security/assessments/2026-11-06-schema.mdfor
how to check whether a running install was affected. bash scripts/setup-selfhosted.shnow needs--yeswhen nobody is at a
terminal. The Supabase CLI asks before it applies migrations, so an unattended
install stops and says so instead of failing halfway. At a terminal, it just asks.
A self-hosted stack that already exists keeps its data where it is.- Apply
20261109000000before putting anyone under a customer key (R452). It
addsuser_data_tables.rows_sealed, off by default. Either order with the build
is safe, and nobody's reads depend on it until they are under a customer key. If
it has not run when such a user writes rows, the write still succeeds, the dataset
stays on the direct path, and the log says[dataset-rows] ... was not marked.
Nothing is encrypted by this yet; the next round needs it. - Apply
20261108000000to seal older approval requests (R450). An approval's
payload cannot be changed once written, so without this the pass leaves older
swarm requests as they are, and the Customer keys card says so. It allows one
change only: the service role, on a swarm step, from text to its sealed form.
New requests and all notifications are sealed without it. It also adds indexes
that keep the pass cheap. - Apply
20261107000000when convenient (R449). It adds partial indexes for
the pass that seals older swarm run traces, so the pass stays cheap once they
are done. Everything works without them, more slowly. - Apply
20261106000000when convenient (R448). It adds a partial index for
the pass that seals older model call traces, so the pass stays cheap once they
are done. Everything works without it, more slowly. - Apply
20261105000000for user keys and customer keys (R438). It adds
user_keys,customer_keys,customer_key_membersand the
destroy_customer_keyfunction, server-only (RLS on, no policies; the
function is the service role's alone). Either order with the build is safe:
until it runs, every credential is written under the deployment's key as
before and the Customer keys card says the migration is needed. After it,
each user gets a key at their next save; run Re-encrypt to current key
once to move what was stored before onto its owners' keys. A deployment
that rotatesPROVIDER_CREDS_SECRETshould run the sweep before removing
the old secret fromPROVIDER_CREDS_SECRET_OLD, as before: it now re-wraps
the user keys instead of re-encrypting every value. - Apply
20261104000000to require a private path for an Iceberg catalog
(R428). It addsiceberg_catalogs.require_private, off for every existing
catalog. Either order with the build is safe: until it runs, every catalog
reads as not requiring the rule, catalogs save as before, and asking for
the rule says which migration to apply. - Check a Bedrock or Claude Platform on AWS private endpoint (R427). The
override must now be the service's own name in the integration's region:
bedrock-runtime.<region>.amazonaws.com(or its FIPS or dual-stack name) or
vpce-….bedrock-runtime.<region>.vpce.amazonaws.com;
aws-external-anthropic.<region>.api.awsor itsvpce-…name. An override
saved before that is any other host now fails its next call with a message
naming the expected name; save the integration again with that name, or
choose Public or Private DNS. Nothing to migrate. - Start this build before applying
20261103000000(R421). The migration
makes swarm run records (swarm_runs, its steps, edges and checkpoints)
read-only to signed-in clients; the previous build wrote canvas runs from
the browser, which the new policy refuses, so in between canvas runs would
go untraced and Cancel would not record. It proves itself on apply. - A secret shared with you now goes only where its owner allows (R419).
Run the migrations (20261102000000addsuser_secrets.allowed_hosts;
either order with the build is safe). Every secret allows no host until its
owner adds some under Secrets → edit → Allowed hosts, so a colleague's
use of a shared secret stops, with an error naming the host to add, until
then. Your own secrets are unaffected. Model providers, the LLM gateway,
n8n, Firecrawl and warehouse connections check a shared secret against the
host they send it to (R420); a notification's webhook URL takes only a
secret you own. An ETL pipeline no longer runs on a connection shared with
its owner: ask the connection's owner to run it, or to give you your own. - The lake's object store is SILO now (R418). MinIO's own images can no
longer be pulled, so Compose and the Kubernetes manifests run SILO, its
maintained fork, under the same service name, volume and settings.
docker compose pull && docker compose up -dreplaces the container in place;
there is nothing to run first and nothing to migrate. - Re-run the installer if your
.envpredates 2026-09-15 (R417). It adds
QDRANT_URLandVECTOR_STOREwhen they are absent, so the bundled Qdrant is
searched; the app copies your existing vectors into it by itself. Keep
VECTOR_STORE="pgvector"to stay on Postgres. - Run the migrations (
npx supabase db push). One adds
ml_experiment_run_write, which experiment logging now calls; until it
exists, logging to a run answers 400. One addsapprovals.resumed_at, which
resuming a swarm run after its approval now claims. One widens the cost
columns to unconstrainednumeric, without rewriting the tables. One (R375)
closes nine service-only database functions to signed-in clients and
requires an ML or MCP API key to name a model or app its owner owns; it
removes any existing key that does not, and fails rather than half-applies
if any of it does not take. One (R377) narrows who may write an approval and
makes an approval's routing immutable; it also proves itself on apply. One
(R383) makesexecution_tracesread-only to signed-in clients — start this
release's build before applying it: the previous build recorded chat and
notebook traces as the signed-in user, which the new policy refuses, so those
traces would go unrecorded in between. It proves itself on apply. One
(20261028000000) lets the encrypted provider store hold every provider the
save form offers — its check admitted only Bedrock, Vertex, Anthropic and
Azure, so OCI Generative AI and Qwen keys could not be saved there — and
Claude Platform on AWS; it also proves itself on apply. One (R389,
20261029000000) limits reading a profile to its owner; it proves itself on
apply, reading a real profile as another user and as its owner. - Set the default model in Admin, not in .env (
20261031000000). A
superadmin sets it under Admin → Developer runtime → AI services →
Default model; until thenopenrouter/google/gemini-2.5-flashis used.
OPENROUTER_DEFAULT_MODEL,SHEETS_ASSIST_MODEL,AI_SQL_DEFAULT_MODEL,
DOCUMENT_VISION_MODELandML_ASSIST_MODELare no longer read; the page
lists any still set. The migration clears the Sheets, AI SQL and vision
model settings where they holdopenrouter/google/gemini-3-flash-preview,
the value the old form saved without anyone choosing it — set them again if
you meant it.
Images
Each is built for linux/amd64 and linux/arm64, signed with Sigstore (keyless, from this repository's release workflow), and carries an SBOM and build provenance.
| Image | Tags | Digest |
|---|---|---|
ghcr.io/agentswarms-fyi/agentswarms |
1.7.0 1.7 latest |
sha256:c8b0f0a8c43d972ddd12874308de3694cabf2cf07d3f81d3ad9873a09c7a22fc |
ghcr.io/agentswarms-fyi/db-migrate |
1.7.0 1.7 latest |
sha256:ba964de6b986efbf4469ebcb9f68879ec52393ef9776260fd8344eca36995aaa |
ghcr.io/agentswarms-fyi/docgen |
1.7.0 1.7 latest |
sha256:1f7ec17a85e1bca67c9b7967573f74853c046787da2e7cbff87325967da91a78 |
ghcr.io/agentswarms-fyi/docker-socket-proxy |
1.7.0 1.7 latest |
sha256:b0f1ab9307165cc8a34db0ba82258f3173aefff1f92db7733f8e0b1978206685 |
ghcr.io/agentswarms-fyi/egress-proxy |
1.7.0 1.7 latest |
sha256:b90f5c46d213a5c9a16e7961980524c0aaa46bd1cf4de493750e6ae147a615ba |
ghcr.io/agentswarms-fyi/js-sandbox |
1.7.0 1.7 latest |
sha256:45c8dad77798e396a4f380895fe8df87112117ca10009653abf00794dfea8f09 |
ghcr.io/agentswarms-fyi/notebook-gateway |
1.7.0 1.7 latest |
sha256:1ba31de76957e3cba8ffcfac45a923a2a3055ef8c14116a863ab638a26e58894 |
ghcr.io/agentswarms-fyi/notebook-runtime |
1.7.0 1.7 latest |
sha256:623f4913003dcb1021b7810dbc5cc0721f53d4bfa3dab026f1ca0c3399ec2235 |
Check a signature before you run an image:
cosign verify ghcr.io/agentswarms-fyi/agentswarms@sha256:c8b0f0a8c43d972ddd12874308de3694cabf2cf07d3f81d3ad9873a09c7a22fc \
--certificate-identity-regexp '^https://github\.com/AgentSwarms-fyi/agentswarms/\.github/workflows/release\.yml@refs/tags/v' \
--certificate-oidc-issuer https://token.actions.githubusercontent.comDatabase
24 new migrations since v1.6.0. They run before the new app starts: npx supabase db push --remote, or the db-migrate image as a one-off task. The previous release runs on the new schema, so rolling the app back one release needs no database change (docs/DEPLOYMENT.md, "Rolling back one release").
20261006120000_ml_experiment_run_write.sql20261006130000_approval_resume_claim.sql20261006140000_cost_full_precision.sql20261025000000_security_rpc_rls_hardening.sql20261026000000_approval_forge_hardening.sql20261027000000_execution_trace_write_lockdown.sql20261028000000_provider_credentials_providers.sql20261029000000_profiles_own_row_only.sql20261030000000_enterprise_ontology.sql20261031000000_default_model.sql20261101000000_embed_keys_owner_check.sql20261102000000_secret_allowed_hosts.sql20261103000000_swarm_run_write_lockdown.sql20261104000000_iceberg_catalog_require_private.sql20261105000000_customer_keys.sql20261106000000_trace_seal_index.sql20261107000000_swarm_run_seal_index.sql20261108000000_approval_seal.sql20261109000000_dataset_rows_sealed.sql20261110000000_catalog_audit_fixes.sql20261111000000_catalog_lake_metadata.sql20261112000000_spark_jobs_server_written.sql20261113000000_retention_in_one_place.sql20261114000000_ecs_sandbox_backend.sql
Removed on purpose (the previous release may not run where it used these):
public.upsert_sample_dataset(20261110000000_catalog_audit_fixes.sql)public.insert_sample_rows(20261110000000_catalog_audit_fixes.sql)