readyagentsdev 1.9.0

ReadyAgents Core 1.9.0

Connector suite: typed contract, SDK, first-party set, catalog CLI.

Typed contract above the tool registry (ConnectorSpec / ConnectorContext),
SDK with Retry-After, pagination and size caps, and a small first-party set
(rest, sql, object_storage, message, ingest). Write-shaped connectors
gate by default; idempotency keys de-dupe retries. readyagents connectors list|show|test is the catalog. Conformance harness fails own-socket /
non-granted-secret / cap-bypass connectors. Catalog is small by design;
governance is the differentiator, not certification.

Also fixed: REST write ops classified from connector_config HTTP method
before the default write gate; message pins destinations to the declaration;
connector HTTP reuses http_get public-IP SSRF pinning including redirects.

See docs/connectors.md and
docs/connector-sdk.md.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.9.0
readyagents connectors list
readyagents connectors show rest
readyagents connectors test rest

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.8.1

ReadyAgents Core 1.8.1

Sovereign patch: honest bundle, resume restore, keyless allowlist.

readyagents bundle fails if pip download cannot collect declared runtime
wheels (a --no-index install needs those wheels, not only the project).
Resuming a sovereign run restores the socket guard and stored allowlist from
the run record even when resume is invoked without --sovereign. Keyless
OpenAI-compat for allowlisted private bases uses the same parsed CLI/env
allowlist as the egress guard (--sovereign-allow 10.0.0.8).

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.8.1
readyagents bundle --out ./offline-wheels
readyagents resume RUN

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.8.0

ReadyAgents Core 1.8.0

Sovereign mode: process-level egress refuse, attest, bundle, doctor.

--sovereign / READYAGENTS_SOVEREIGN=1 refuses non-loopback egress at the
process socket boundary for the run (model calls, tools, packs, threads).
Loopback is allowed; private endpoints are explicit (--sovereign-allow).
readyagents attest emits a residency document that marks MCP stdio
network_uncontrolled: true and does not claim legal compliance.
readyagents bundle writes wheels plus checksums for
pip install --no-index --find-links. readyagents doctor reports whether
sovereign would succeed here and loopback model presence, never secret values.
Keyless local OpenAI-compatible endpoints are allowed. In-process is not an OS
sandbox. See docs/sovereign.md and
docs/local-models.md.

Also fixed: scoped delegation matches the gate's approver_roles; JSON
run-record reads retry Windows sharing violations; sequential HITL resume waits
out a prior in-flight executor when the next gate is already paused.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.8.0
readyagents run examples/calc_pipeline.yaml --sovereign
readyagents attest RUN --json
readyagents bundle --out ./offline-wheels
readyagents doctor

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.7.0

ReadyAgents Core 1.7.0

Enterprise HITL: quorum, roles, deadlines, delegation.

Opt-in quorum (approvals_required), distinct actors, deny_actor (including
$initiator), role routing (approver_roles / require: any|all), lazy
deadlines (expires_in with on_expire: reject | escalate | fail — approve
is refused at validation), time-bounded revocable single-hop delegation
(readyagents delegate / delegations list|revoke), require_reason and
override recording, file / command / webhook notify channels, and
readyagents approvals list. Core starts no timer; expiry is evaluated on
resume, decide, and status query. A gate with none of the new fields is
unchanged.

Also: MCP decide can resume a recorded decision that is still paused, so
sequential approval resumes do not stall after the first decide.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.7.0
readyagents approvals list
readyagents delegate --help

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.6.0

ReadyAgents Core 1.6.0

Supply-chain trust: digests, detached signatures, keyring, lock, SBOM.

Canonical SHA-256 digests (algorithm v1) for workflows including every resolved
include, pack file bytes, and MCP advertised tool surfaces. readyagents sign
/ verify write a detached Ed25519 signature beside the artifact that binds
digest and kind. readyagents trust add|list|remove manages a local publisher
keyring under $READYAGENTS_HOME. --require-signed and policy require_signed
refuse unsigned or untrusted artifacts before a pack is imported.
readyagents lock and --frozen pin digests; readyagents sbom emits a
deterministic CycloneDX-shaped inventory. Optional sign extra is not in all.
No default-trusted key. Signing proves origin, not safety.

Also: --require-signed executes the digested workflow/include buffers (not a
later re-read), pack lock pins are checked before import, templated include
paths are refused under --require-signed / --frozen, and
on_lock_mismatch: gate persists a paused run that can be resumed.

Packs are waitlisted, not for sale.

Try it

pip install 'readyagentsdev[sign]==1.6.0'
readyagents trust list
readyagents verify examples/calc_pipeline.yaml
readyagents sbom examples/calc_pipeline.yaml

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e '.[sign]'.

readyagentsdev 1.5.1

ReadyAgents Core 1.5.1

Identity skeptic fixes: outbound notify JWTs and brokered secret isolation.

Workload sign_assertion is attached to outbound pause-notify (post_json)
when configured so a fixture peer can verify the JWT against the public key.
Credential grants are delivered via a thread-local mapping, not shared
os.environ, so parallel tool branches cannot see a sibling's secret.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.5.1
readyagents identity whoami

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.5.0

ReadyAgents Core 1.5.0

Agent identity and credential brokering.

Approvers may present an OIDC/JWT assertion verified against a local trust-anchor
file (--token-file, --trust-anchors / READYAGENTS_TRUST_ANCHORS). Verification
uses the optional jwt extra (not in all): signature, issuer, audience, expiry,
skew. alg: none, algorithm confusion, and unknown kid are refused. Fail closed
on a missing/malformed/unreadable anchor when a token is presented. Claims map to
--actor / RBAC roles; replay of the same token on a gate is refused. Signed
(HMAC of the decision body) and identified (verified subject) stay separate;
--actor NAME remains the default.

Optional readyagents.credentials.yaml grants named secrets per tool at the
dispatch seam; non-granted tools cannot read them from os.environ during the
call. credential_kind is static when the provider cannot mint.

Also: readyagents identity verify|whoami|trust (workload whoami prints a
fingerprint, never the private key), plus TokenOps spend-meter/ledger fixes for
parallel reserves and unpriced cost_micros.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.5.0
readyagents identity trust
readyagents identity whoami

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.4.0

ReadyAgents Core 1.4.0

TokenOps: estimate, spend caps, ledger, and runaway guards.

Versioned, overridable model price table (READYAGENTS_PRICES); unknown models
are explicitly unpriced, never a silent zero. readyagents run PATH --estimate
walks the engine's routing with no execute and no network and prints a range
with assumptions. --max-spend / --max-tokens are consulted before each model
call; parallel branches share one meter; a resumed run continues the same
budget. --label KEY=VALUE is stored on the run and in an append-only
hash-chained spend ledger (readyagents spend). Cache hits/misses/savings
appear on the run record, runs report, and the ledger. Runaway guards
(--max-model-calls, --max-run-tool-rounds, --max-wall-seconds, workflow
runaway:) raise RunawayGuard, distinct from BudgetExceeded and
CircuitOpen. Optional tokenizer extra is not in all. The provider invoice
is authoritative. Without the new flags, behaviour is unchanged.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.4.0
readyagents run examples/calc_pipeline.yaml --estimate
readyagents spend

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.3.0

ReadyAgents Core 1.3.0

Traceability evidence: hash-chained audit, evidence packs, retention, observers.

Audit JSONL is hash-chained (seq, prev_hash, entry_hash) — tamper-evident,
not tamper-proof. readyagents audit verify reports unchained ranges and the
first break. readyagents evidence RUN_ID writes a hash-manifested local pack
(machine JSON, self-contained HTML, decisions projection, audit slice, workflow
source, Mermaid graph); the pack may contain prompts and outputs.
readyagents graph PATH is deterministic and injection-safe. Configurable
retention (READYAGENTS_RETENTION_DAYS, default 180) makes runs gc refuse
in-window records unless --override-retention (audited). Pack observer seam
(register_observers) is backward-compatible. Optional content-free otel
extra is not in all, starts no collector on import, and is off unless
READYAGENTS_OTEL=1. Docs claim evidence, never compliance or certification.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.3.0
readyagents run examples/calc_pipeline.yaml
readyagents audit verify
readyagents evidence RUN_ID
readyagents graph examples/graph_complex.yaml

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..

readyagentsdev 1.2.1

ReadyAgents Core 1.2.1

Firewall skeptic fixes for taint, policy, and MCP pins.

Agent tool-calls inherit taint from the calling prompt/system so on_tainted
applies when the model emits literal arguments. Foreach copies parent
provenance and marks item/index untrusted when the items source is
untrusted. The resolved policy path and MCP pins persist across resume/decide
and later runs, so omitting --policy on resume cannot fail-open a gate.
nodes.<id>.require_approval: only approve proceeds; reject is a policy deny.

Packs are waitlisted, not for sale.

Try it

pip install readyagentsdev==1.2.1
readyagents policy check examples/readyagents.policy.yaml
readyagents run examples/policy_gated.yaml --policy examples/readyagents.policy.yaml

Or clone https://github.com/readyagentsdev/readyagents-core and pip install -e ..