Okto Nexus

August 23, 2026 · View on GitHub

Local-first coordination for teams of AI agents.

Okto Nexus is an MCP server and operator hub for agents working in the same repository. It gives them durable identities, presence, messages, inboxes, handoffs, artifacts, an event log, governance controls, and a live dashboard without requiring a cloud broker.

okto-nexus serve exposes the complete hub on one port:

/mcp — MCP over streamable HTTP, authenticated as an agent;
/api/v1 — operator REST APIs, read-only monitor endpoints, and SSE;
/ — the bundled React dashboard.

The classic stdio transport remains supported. Coordination records live in one SQLite database in WAL mode. Optional metrics, referenced workspace files, and the derived shared.md view live outside that database.

Release fact	Value
Package	`okto-nexus 0.1.4`
Python	`>=3.11`
MCP surface	43 tools by default; 46 with memory enabled
MCP resources	12 versioned reference resources
MCP prompts	0
Surface revision	32
Database schema	26 migrations, 34 tables
Storage	local SQLite/WAL

PyPI: pypi.org/project/okto-nexus
Source: github.com/OktoLabsAI/okto-nexus

Why Nexus
Install
Start the hub
Connect an MCP client
Agent pre-flight
Architecture
HTTP surfaces and authentication
Dashboard
Coordination model
MCP surface
Configuration
Operations
Data model and migrations
Errors
Development
Contributing
Troubleshooting
Security and limitations
Release notes
License

Why Nexus

One coordination space per project. An absolute project_root is canonicalized and hashed into a deterministic workspace_id. Every client that resolves the same real path joins the same workspace.
Durable delivery. Messages fan out into per-recipient inbox lanes with leases, redelivery, acknowledgements, delivery status, and optional read receipts.
Single-winner work dispatch. Handoffs support atomic claim, leases, rejection, cancellation, optional verification, and optional dependency graphs.
Explicit identity and presence. Operators create identities and API keys; agents open sessions and heartbeat to remain present.
Targeted routing. Direct, capability, role, tag, broadcast, mixed, and direct-with-fallback strategies share one validated grammar.
Governed communication. Permissions, communication scopes, versioned policies, quotas, guardrails, groups, and human approval can restrict writes without exposing the control plane to agents.
Observable by design. Monotonic event IDs, cursor reads, long-poll, replay export, SSE, health aggregates, and a live dashboard expose what the team is doing.
Local-first and fail-closed. Configuration, target grammars, catalogs, workspace paths, API keys, and state transitions are validated before writes.
Token-aware MCP docs. First-use guidance remains resident; deeper reference material is available through versioned MCP resources on demand.

Install

The recommended install includes the HTTP hub, dashboard, local embedding provider, and tokenizer:

uv tool install "okto-nexus[serve]"
okto-nexus serve

Equivalent with pipx:

pipx install "okto-nexus[serve]"

Available extras:

Extra	Includes	Use it when
none	stdio MCP core	You only need a lightweight local stdio server
`serve-lite`	FastAPI, Uvicorn, dashboard	You need HTTP without Torch/model dependencies
`embeddings`	sentence-transformers	You want the local embedding provider separately
`serve`	HTTP stack, embeddings, tokenizer	You want the complete supported hub
`dev`	pytest, FastAPI, Uvicorn, httpx	You are developing or testing Nexus

The published wheel and sdist contain the compiled dashboard. Node.js is only needed when rebuilding the frontend from source.

From a checkout:

git clone https://github.com/OktoLabsAI/okto-nexus.git
cd okto-nexus
uv sync --extra dev

# Complete HTTP build:
uv sync --extra serve --extra dev

Start the hub

okto-nexus serve

Defaults:

dashboard: http://127.0.0.1:8202/;
MCP: http://127.0.0.1:8202/mcp;
data directory: ~/.okto_nexus;
database: ~/.okto_nexus/nexus.db;
initial workspace context: the current directory. The dashboard keeps a saved selection when present and otherwise may open the all-workspaces view.

Useful variants:

okto-nexus serve --project-root /absolute/path/to/project
okto-nexus serve --host 0.0.0.0 --port 8202
okto-nexus serve --trust-mode strict
okto-nexus serve --embedding-mode local

On first use, open Agents → New agent in the dashboard. Create one identity per participant and copy its nxs_... key immediately: Nexus stores only the hash and shows plaintext only at creation or regeneration. Capability and tag values must first exist in Registry before an identity can use them.

Connect an MCP client

The dashboard generates snippets for Claude Code, Claude Desktop, Codex, Cursor, VS Code, Windsurf, and Cline. The generic streamable-HTTP URL is:

http://127.0.0.1:8202/mcp?api_key=nxs_REPLACE_ME

Credential extraction order is query api_key, x-api-key header, then Authorization: Bearer. Treat client configuration containing a query key as a secret.

Examples:

claude mcp add -t http okto-nexus \
  "http://127.0.0.1:8202/mcp?api_key=nxs_REPLACE_ME"

codex mcp add okto-nexus \
  --url "http://127.0.0.1:8202/mcp?api_key=nxs_REPLACE_ME"

Generic JSON:

{
  "mcpServers": {
    "okto-nexus": {
      "url": "http://127.0.0.1:8202/mcp?api_key=nxs_REPLACE_ME"
    }
  }
}

Stdio

Run okto-nexus without a subcommand for stdio:

{
  "mcpServers": {
    "okto-nexus": {
      "command": "okto-nexus",
      "args": [],
      "env": {
        "OKTO_NEXUS_HOME": "/absolute/path/to/nexus-home"
      }
    }
  }
}

Without OKTO_NEXUS_API_KEY, stdio preserves the cooperative anonymous model. Set that variable to an active nxs_... key to bind the process to the same authenticated identity rules as HTTP. An invalid configured key fails closed.

Agent pre-flight

Every authenticated agent should do this on its first turn:

Call agent_whoami(), use the returned agent_id consistently, and treat its role and communication.content as the default operating contract.
Call workspace_resolve(project_root=<absolute cwd>).
Call session_open(agent_id=<you>, workspace_id=<resolved id>) and retain the returned session_id and one-time session_secret.
Check inbox_count(agent_id=<you>); pull and acknowledge backlog.
Anchor monitoring with event_cursor(project_root=..., agent_id=<you>, stream="workspace").

The full procedure is available at okto-nexus://reference/preflight.

The role guides responsibilities, operating perspective, and decision boundaries. The communication block guides tone, format, language, verbosity, structure, and agent-to-agent as well as user-facing communication. Follow both unless the user explicitly directs otherwise for the current task or interaction. That task-scoped override does not modify the Nexus profile or bypass permissions, policies, guardrails, approvals, communication scope, safety rules, or higher-priority host instructions. Capabilities are routing claims, not authorization or persona.

Direct message

{
  "project_root": "/absolute/path/to/project",
  "from_agent_id": "researcher",
  "subject": "API findings",
  "body": "The endpoint is idempotent; details are attached.",
  "target": {
    "strategy": "direct",
    "agent_id": "implementer"
  },
  "from_session_id": "ses_...",
  "session_secret": "..."
}

The response names the resolved recipients and delivery count. The recipient uses:

inbox_count(agent_id="implementer")
inbox_pull(agent_id="implementer", session_id="ses_...", session_secret="...")
inbox_ack(agent_id="implementer", message_ids=[...],
          session_id="ses_...", session_secret="...")

Messages are delivered through the inbox. event_get and event_wait are observability tools, not delivery.

Long-poll

event_wait is a snapshot when timeout_seconds is omitted, null, or 0. Long-poll is explicit:

event_wait(
  project_root="/absolute/path/to/project",
  agent_id="researcher",
  stream="workspace",
  cursor=123,
  timeout_seconds=25,
  profile="summary"
)

Always continue from next_cursor. The waiter uses SQLite PRAGMA data_version plus bounded sleep polling; HTTP runs the blocking wait in a worker thread so it does not block the shared event loop.

Architecture

MCP stdio             MCP HTTP              REST / SSE / SPA        CLI
    \                     |                         |                 /
     +---------------- inbound adapters and transport auth ----------------+
                                      |
                             application services
       identity · messages · inbox · handoffs · events · artifacts
       permissions · policies · approvals · guardrails · memory · health
                                      |
                         domain models and pure rules
                                      |
      +------------------------- outbound ports ---------------------------+
      | SQLite repositories | files/shared.md | waiter | telemetry | embed |
      +--------------------------------------------------------------------+

bootstrap() resolves configuration, creates the store, applies migrations, wires repositories, telemetry, embeddings, and approval execution, then seeds the reserved operator, backfills the capability catalog, and creates built-in permission/communication presets. create_server() lazily imports FastMCP and registers the effective tools and resources. serve wraps the same composition in FastAPI/Uvicorn; tail and admin are separate CLI adapters.

Important boundaries:

domain code contains state machines, routing, IDs, and invariants;
application services own the core coordination use cases; operator CRUD/maintenance routes may drive repositories and units of work directly;
inbound adapters translate MCP, HTTP, SSE, and CLI calls;
outbound adapters implement SQLite, files, telemetry, tokenization, embeddings, and waiting;
coordination truth is durable in SQLite; bounded process caches are implementation details, not authoritative state.

HTTP surfaces and authentication

Surface	Loopback bind	Non-loopback bind
SPA shell/assets, `/healthz`, info, license	Public	Public
REST data/control plane	Keyless operator trust	Active `nxs_` key required
MCP `/mcp`	Active `nxs_` key required	Active `nxs_` key required
EPT monitor endpoints	Scoped `nxsept_` accepted	Scoped `nxsept_` accepted

MCP-over-HTTP connections always represent an agent; stdio may use the cooperative anonymous mode. The dashboard/REST loopback trust path represents the local operator. Browser-origin checks protect mutating operator routes, and binding beyond loopback removes keyless REST trust.

On a non-loopback bind, use the reserved operator identity's key for the dashboard/control plane. Participant keys authenticate requests but operator-only routes return PERMISSION_DENIED. When a store has no keys at all, startup creates the operator key and prints its plaintext once.

Permanent agent keys can authenticate REST and MCP, but helper monitors should receive only a short-lived ephemeral poll token (nxsept_...). EPTs are bound to the issuing session, agent, and workspace and are accepted only as Authorization: Bearer nxsept_...; query api_key and x-api-key are rejected for this token type. The bearer is valid only on:

GET /api/v1/events and GET /api/v1/events/cursor;
GET /api/v1/inbox/count and GET /api/v1/inbox/peek.

They cannot call MCP or mutate state.

Dashboard

The bundled dashboard provides:

Graph — toggle between detailed agent cards and compact activity-sized circles, with profile colours, live presence status badges, recent message flow, unread traffic, open handoffs, and claimed relationships;
Messages — inbox lanes, peer conversations, undelivered targeting outcomes, receipts, and optional semantic search;
Handoffs — a six-column Kanban including VERIFYING, claim details, dependency state, verification, cancellation, and results;
Events — filtered event history, trace navigation, and live SSE updates;
Memory — durable memory browse/search/curation when feature_memory is enabled;
Workspaces — sessions, analytics, and coordination health;
Agents — identities, keys, activation, roles, capabilities, metadata, colors, permissions, tags, inbound/outbound audiences, communication style, and steering;
Registry — operator-managed capability and tag vocabularies;
Policies — versioned policies and per-agent bindings;
Guardrails — groups, versioned content rules, assignments, and scrubbed denial audit;
Communication — versioned communication presets and bindings;
Approvals — pending and decided human-in-the-loop actions;
Settings — runtime-manageable settings, feature flags, retention, and database maintenance; metrics use their own header-menu panel.

Semantic search requires embedding_mode=stub or local. off returns EMBEDDINGS_UNAVAILABLE on the REST search endpoint. The local provider needs the embeddings extra; stub is deterministic and is intended for tests or demonstrations.

Coordination model

Workspaces, agents, and sessions

Agents are global identities; workspaces represent canonical project roots.
Most coordination tools accept project_root.
session_open and shared_md_render consume a resolved workspace_id.
An authenticated agent can update only its own profile with agent_register, subject to identity.update_profile and identity.update_capabilities. Operators create identities.
Authenticated discovery is reachability-scoped. agent_list and agent_get hide unreachable peers; capability_list returns the complete catalog but filters owner identities.
workspace_list is permission-gated; absolute paths require a separate permission.
Cross-workspace errors are intentionally operation-specific: WORKSPACE_MISMATCH for ownership guards, NOT_FOUND for hidden artifact or memory reads, and DEPENDENCY_NOT_FOUND for dependency creation.

Presence is explicit. A session is considered present while its heartbeat is within presence_ttl_seconds. A trust-sensitive write advances the heartbeat only when it authenticates with that session's credentials; in trust_mode=open, a credential-free write advances no session. Read-only tools do not heartbeat. Call session_heartbeat during long read-only or idle periods and session_close when finished.

Messages and inboxes

message_create persists one message and resolves recipients at send time. Each recipient gets a durable delivery row in its global inbox. The lanes are:

Lane	Meaning
`unread`	Available to pull
`delivered`	Pulled and protected by an in-flight lease
`read`	Acknowledged
`parked`	Dead-lettered after exhausting delivery claims; not redelivered automatically

An expired delivered lease becomes pullable again. Delivery is therefore at-least-once until acknowledgement. message_status lets the sender inspect each recipient's lane. Every pull/redelivery emits message.delivered; ack emits message.read. By default, ack also sends one synthetic read-receipt message to the sender; receipts do not recursively create receipts.

Channels are organizational labels, not ACLs or delivery mechanisms. Access still intersects with permissions, policy, guardrails, and communication reachability. Message retention can remove aged messages and their deliveries, including unread or in-flight rows, so durability is bounded by configured retention and explicit database reset.

Communication intent

Choose the coordination mechanism by intended outcome: if another agent is expected to perform work or produce a deliverable, create a handoff. If the recipient only needs to know something or reply, use a message.

Handoff — executable work. Any request to execute, investigate, change, build, test, review, validate, or otherwise produce a deliverable must use handoff_create; a direct message must not be its sole record. Target the intended assignee directly when known, or use a capability, role, tag, mixed, broadcast, or direct-with-fallback target when the first eligible claimant should own the work. The handoff is the canonical, operator-visible record of ownership, lifecycle, result, and verification. Include the objective, context, scope, constraints, and expected deliverable; add acceptance criteria, dependencies, and a verifier when applicable and the corresponding feature_verification / feature_dag flag is enabled. Those fields are rejected while their feature is off.
Broadcast message — shared alignment. Use it for shared context, decisions, announcements, discoveries, risk or blocker alerts, and general alignment across every selected reachable recipient. It informs recipients but assigns no owner and creates no task lifecycle. If anyone is expected to act, create one or more handoffs.
Direct message — conversation and informal coordination. Use it for status checks, questions, clarifications, acknowledgements, focused context exchange, and informal coordination. It may discuss an existing handoff, but new executable work requires a handoff; reference that handoff in the conversation. Use handoff_get for canonical lifecycle state and direct messages for contextual updates or blocker explanations.

A broadcast message is informational fan-out to many recipients. A handoff with a broadcast target is a claim pool for one executor. If several agents must produce independent results, create separate handoffs.

Routing

There are seven strategies:

Strategy	Descriptor	Notes
`direct`	`{"strategy":"direct","agent_id":"a"}`	One named identity
`capability`	`{"strategy":"capability","capability":"review"}`	One or any of several registered capabilities
`role`	`{"strategy":"role","role":"reviewer"}`	Exact role match
`tag`	`{"strategy":"tag","selector":{"team":["platform"]}}`	Registered tag selector
`broadcast`	`{"strategy":"broadcast"}`	Present workspace agents for messages; globally registered eligible agents for handoffs
`mixed`	`{"strategy":"mixed","rules":[...]}`	Non-empty union of non-broadcast rules
`direct_with_fallback`	direct plus `fallback_after_seconds` and optional `fallback`	Handoffs only

Messages support direct, capability, role, tag, broadcast, and mixed. Omitting a message target means broadcast. Handoffs require an explicit target and additionally support direct-with-fallback.

Tag selectors use AND across keys and OR across values. Rich In/NotIn/Exists/DoesNotExist expressions are also supported. Capability and tag names fail closed against operator-managed catalogs.

Target resolution is intersected with presence where applicable and with the caller's effective communication reach/audience. Separate enforcement layers then allow, deny, limit, or intercept the write:

per-agent permissions and recipient/rate limits;
versioned policy action rules and quotas;
content guardrails;
optional HITL approval interception.

A channel itself adds no ACL. Visibility controls who may see an item; eligibility controls who may claim it.

Event log

Events are immutable after insertion during normal operation. event_id is globally monotonic and never reused. Retention may delete old rows, so retained history can contain gaps.

Streams are workspace, agent, and handoff. Supported filters are type, agent_id, task_id, handoff_id, and trace_id. Authenticated event reads require events.read and omit actors outside the caller's communication reach, except the caller's own and system events.

Response profiles:

Profile	Behavior
`default`	Safe trim: preserves contextual fields while removing empty or duplicated data; oversized event payloads can yield a follow-up hint
`summary`	Aggressive projection: omits heavy bodies/payloads and returns follow-up hints
`full`	Raw debugging escape hatch

event_get is non-blocking. event_cursor returns the current end in O(1). event_wait long-polls only when timeout_seconds > 0. Dashboard SSE already provides operator UI updates; agent monitoring remains MCP polling/long-poll or the EPT read-only REST plane.

Handoffs

OPEN      -> CLAIMED
OPEN      -> REJECTED | CANCELLED
CLAIMED   -> COMPLETED
CLAIMED   -> VERIFYING        when acceptance criteria exist
CLAIMED   -> REJECTED         claimant rejects
CLAIMED   -> OPEN             lease expires
VERIFYING -> COMPLETED        verifier passes
VERIFYING -> CLAIMED          verifier fails; executor reworks

Only one agent wins handoff_claim. A claim returns the confidential payload. Available-list responses omit payload; handoff_get includes it only for the claimant. Events and synthetic notifications never carry the payload.

With feature_verification=true, creation may include acceptance_criteria and verify_by. The executor cannot verify its own result. A failed verdict persists feedback, returns the handoff to CLAIMED, and renews the lease.

With feature_dag=true, creation may include depends_on. The persisted status remains OPEN, but blockedness is derived. Blocked items are excluded from handoff_list_available and claim returns DEPENDENCY_NOT_MET until every dependency is COMPLETED.

Artifacts and `shared.md`

Artifacts are either inline text/json/markdown or a workspace-contained file reference. Path containment is checked after canonical resolution; escapes return PATH_OUTSIDE_WORKSPACE. Inline content is bounded by max_inline_bytes.

The publisher's effective outbound audience is frozen on the artifact. The same audience controls artifact.created visibility and artifact_get. Unauthorized, cross-workspace, and missing reads all return NOT_FOUND.

shared_md_render atomically rewrites a derived workspace shared.md with four fixed sections: relevant agents/sessions, open tasks, open handoffs, and recent events. It never becomes the source of truth and never renders handoff payloads. Authenticated callers need shared_md.render.

Permissions, policies, guardrails, and approvals

The operator control plane manages:

permission presets and per-agent effective permissions;
capability/tag catalogs and communication scopes;
versioned attachable policies with deny-overrides and quota windows;
versioned communication guidance returned privately by agent_whoami;
agent groups and versioned guardrails assigned by scope and priority;
one-shot HITL approval decisions and operator steering.

Guardrail denials persist scrubbed metadata, not rejected raw content. When feature_hitl=true and a policy requires approval, message_create or handoff_create may return status: "pending_approval". Do not resend the action; watch the returned approval ID.

Opt-in features

All seven flags default to off:

Flag	Effect
`feature_trace`	Accept and project `trace_id` correlation
`feature_hitl`	Enable new `require_approval` interception
`feature_verification`	Enable handoff acceptance criteria and verdicts
`feature_dag`	Enable handoff dependencies
`feature_memory`	Register three memory MCP tools and show Memory UI; restart required
`feature_health`	Enable `coordination_health` MCP execution
`feature_replay`	Enable REST replay export

Nuances:

approval history and decisions remain available when interception is off;
disabling verification or DAG blocks new contracts/dependencies, while already persisted workflow state remains enforceable and decidable;
workspace health REST remains available when the MCP health flag is off;
CLI replay export is operator-shell access and is not gated by feature_replay;
memory REST supports operator curation independently, but MCP memory tools are registered only when the flag is on at startup.

MCP surface

The default server exposes 43 tools: 42 across tool modules plus nexus_info. Enabling feature_memory at startup adds three tools, for 46. Both transports expose the same effective tool/resource surface for the same configuration.

Area	Tools
Metadata	`nexus_info`
Identity/workspace	`workspace_resolve`, `agent_register`, `agent_whoami`, `session_open`, `session_heartbeat`, `session_close`, `workspace_list`, `agent_list`, `agent_get`, `capability_list`
Events	`event_get`, `event_cursor`, `event_wait`
Messages/channels	`message_create`, `channel_create`, `channel_list`, `message_get`, `message_list`, `message_wait`
Inbox	`inbox_pull`, `inbox_ack`, `inbox_extend`, `inbox_peek`, `inbox_count`, `inbox_history`, `message_status`
Handoffs	`handoff_create`, `handoff_list_available`, `handoff_claim`, `handoff_complete`, `handoff_verify`, `handoff_reject`, `handoff_cancel`, `handoff_get`
Artifacts	`artifact_put`, `artifact_get`
Derived view	`shared_md_render`
Health	`coordination_health`
Catalog	`tag_list`
Ephemeral monitor tokens	`poll_token_issue`, `poll_token_renew`, `poll_token_revoke`
Optional memory	`memory_put`, `memory_get`, `memory_search`

message_get, message_list, and message_wait are intentional migration shims. They return MIGRATED with replacements in the inbox/event surface instead of failing as unknown tools.

coordination_health stays registered but returns VALIDATION_ERROR while feature_health is off. memory_put/get/search are absent until feature_memory is enabled and the server restarts.

Session credentials are trust-sensitive on:

message_create;
handoff_claim/complete/verify/reject/cancel;
inbox_pull/ack/extend;
memory_put when published.

In trust_mode=open they are optional but validated if supplied. In trust_mode=strict they are required. poll_token_issue/renew/revoke always require a valid session_id and session_secret in both modes.

Versioned MCP resources

URI	Version
`okto-nexus://reference/preflight`	4
`okto-nexus://reference/communication`	3
`okto-nexus://reference/monitoring`	5
`okto-nexus://reference/target-grammar`	5
`okto-nexus://reference/tool-docs/messages`	3
`okto-nexus://reference/tool-docs/inbox`	2
`okto-nexus://reference/tool-docs/events`	2
`okto-nexus://reference/tool-docs/handoff`	4
`okto-nexus://reference/tool-docs/identity`	5
`okto-nexus://reference/tool-docs/artifacts`	2
`okto-nexus://reference/governance`	2
`okto-nexus://reference/hitl`	2

nexus_info reports package/schema/surface versions, the URI-to-version map, and effective feature flags. Use that live metadata instead of assuming a cached surface.

Response envelope

Successful tools return:

{
  "ok": true,
  "data": {}
}

Failures return:

{
  "ok": false,
  "error": {
    "code": "VALIDATION_ERROR",
    "message": "Human-readable explanation"
  }
}

Transient SQLite lock/busy failures use DB_ERROR with details.retryable=true.

Resident token footprint

For the 0.1.4 default surface:

Component	Characters
Server instructions	3,992
Tool docstrings	6,563
Parameter schemas/descriptions	15,720
Cuttable resident surface	26,275 (~6,568 tokens)
Total measured surface	37,345 (~9,336 tokens)

Deep explanations live in resources so clients load them only when needed. The current measured cuttable reduction against the frozen baseline is about 44.3%.

Configuration

For serve settings managed by the runtime catalog, effective precedence is:

CLI flag > environment variable > stored dashboard override > default

Within the core/stdio bootstrap there is no stored layer, so precedence is CLI > env > default. Unknown flags, missing values, invalid enums, and out-of-range numbers fail closed with CONFIG_ERROR. Boolean CLI flags take an explicit value such as --feature-trace true.

Core runtime

Environment	CLI	Default	Notes
`OKTO_NEXUS_HOME`	`--home`	`~/.okto_nexus`	Runtime data directory
`OKTO_NEXUS_DB_PATH`	`--db-path`	`{home}/nexus.db`	SQLite database
`OKTO_NEXUS_BUSY_TIMEOUT_MS`	`--busy-timeout-ms`	5000	Minimum 0
`OKTO_NEXUS_POLL_INTERVAL_MS`	`--poll-interval-ms`	200	Waiter interval; minimum 1
`OKTO_NEXUS_MAX_WAIT_TIMEOUT_SECONDS`	`--max-wait-timeout-seconds`	30	Server wait ceiling; minimum 0
`OKTO_NEXUS_HANDOFF_LEASE_TTL_SECONDS`	`--handoff-lease-ttl-seconds`	300	Minimum 1
`OKTO_NEXUS_MAX_INLINE_BYTES`	`--max-inline-bytes`	65536	Inline artifact/content limit
`OKTO_NEXUS_INBOX_LEASE_TTL_SECONDS`	`--inbox-lease-ttl-seconds`	300	Minimum 1
`OKTO_NEXUS_SESSION_STALE_TTL_SECONDS`	`--session-stale-ttl-seconds`	60	Derived stale threshold
`OKTO_NEXUS_PRESENCE_TTL_SECONDS`	`--presence-ttl-seconds`	1800	Broadcast/tag presence window
`OKTO_NEXUS_SESSION_REAP_SECONDS`	`--session-reap-seconds`	86400	Opportunistic stale close
`OKTO_NEXUS_MAX_SHARED_MD_EVENTS`	`--max-shared-md-events`	1000	Render ceiling
`OKTO_NEXUS_MAX_EVENT_LIMIT`	`--max-event-limit`	1000	Event page ceiling
`OKTO_NEXUS_POLL_TOKEN_TTL_SECONDS`	`--poll-token-ttl-seconds`	3600	Minimum 60
`OKTO_NEXUS_TRUST_MODE`	`--trust-mode`	`open`	`open` or `strict`
`OKTO_NEXUS_EMBEDDING_MODE`	`--embedding-mode`	`off`	`off`, `stub`, or `local`
`OKTO_NEXUS_INBOX_READ_RECEIPTS`	`--inbox-read-receipts`	`true`	Sender inbox receipts
`OKTO_NEXUS_EXPOSE_WORKSPACE_PATH`	`--expose-workspace-path`	`false`	Operator REST/dashboard path disclosure
`OKTO_NEXUS_AUTO_PRUNE_ON_START`	`--auto-prune-on-start`	`false`	One bounded best-effort startup pass

Retention

Environment	CLI	Default	Minimum
`OKTO_NEXUS_RETENTION_EVENTS_KEEP_DAYS`	`--retention-events-keep-days`	30	0
`OKTO_NEXUS_RETENTION_READ_DELIVERIES_KEEP_DAYS`	`--retention-read-deliveries-keep-days`	14	0
`OKTO_NEXUS_RETENTION_CLOSED_SESSIONS_KEEP_DAYS`	`--retention-closed-sessions-keep-days`	7	0
`OKTO_NEXUS_RETENTION_MESSAGES_KEEP_DAYS`	`--retention-messages-keep-days`	30	7

Pruning removes aged events, read deliveries, closed sessions, and messages older than the message window. Message retention is pure-age: it can remove unread, in-flight, or parked messages and cascades to their deliveries and embeddings. Handoffs and other non-message live rows are not age-pruned.

Metrics

Environment	CLI	Default	Notes
`OKTO_NEXUS_METRICS_MODE`	`--metrics-mode`	`disabled`	`disabled`, `local_only`, `anonymous_beacon`
`OKTO_NEXUS_METRICS_DIR`	`--metrics-dir`	`{home}/metrics`	Local telemetry JSONL/state
`OKTO_NEXUS_METRICS_BEACON_URL`	`--metrics-beacon-url`	`https://nexus-metrics.oktolabs.ai`	Used only in beacon mode
`OKTO_NEXUS_METRICS_RETENTION_DAYS`	`--metrics-retention-days`	30	Minimum 0
`OKTO_NEXUS_METRICS_PUBLISH_INTERVAL_SECONDS`	`--metrics-publish-interval-seconds`	3600	Minimum 60

Metrics are opt-in. Local mode stores bounded per-event metadata; beacon mode publishes aggregate hourly counts only. Message bodies, prompts, workspace file paths, coordination IDs, keys, tokens, URLs, and stack traces are excluded. Local telemetry JSONL is not currently pruned automatically; operators must manage those files even though metrics_retention_days is validated and exposed in configuration.

Feature flags

Environment	CLI	Default
`OKTO_NEXUS_FEATURE_TRACE`	`--feature-trace`	`false`
`OKTO_NEXUS_FEATURE_HITL`	`--feature-hitl`	`false`
`OKTO_NEXUS_FEATURE_VERIFICATION`	`--feature-verification`	`false`
`OKTO_NEXUS_FEATURE_DAG`	`--feature-dag`	`false`
`OKTO_NEXUS_FEATURE_MEMORY`	`--feature-memory`	`false`
`OKTO_NEXUS_FEATURE_HEALTH`	`--feature-health`	`false`
`OKTO_NEXUS_FEATURE_REPLAY`	`--feature-replay`	`false`

feature_memory changes tool registration and requires restart/reconnect. The other flags gate live behavior.

Serve-only and transport-specific settings

Environment	CLI	Default	Scope
`OKTO_NEXUS_PORT`	`--port`	8202	`serve`
`OKTO_NEXUS_HOST`	`--host`	`127.0.0.1`	`serve`
`OKTO_NEXUS_LOG_LEVEL`	`--log-level`	`warning`	`critical` through `trace`
—	`--project-root`	`.`	Initial dashboard workspace
`OKTO_NEXUS_NO_BANNER`	—	unset	Suppress serve banner
`OKTO_NEXUS_API_KEY`	—	unset	Optional stdio authenticated identity

Operations

CLI commands

Command	Purpose
`okto-nexus serve`	Start MCP HTTP, REST, SSE, and dashboard
`okto-nexus`	Start MCP over stdio
`okto-nexus tail`	Operator NDJSON follower over the event service
`okto-nexus admin prune`	Enforce retention, optionally vacuum
`okto-nexus admin issue-keys`	Add keys to legacy keyless identities
`okto-nexus admin export`	Export a workspace replay stream as NDJSON

Use --help on every command for the full argument grammar.

Tail

okto-nexus tail \
  --project-root /absolute/path/to/project \
  --agent-id observer \
  --stream workspace \
  --from latest

tail applies per-agent event visibility. An optional --cursor-file belongs to that consumer only; corrupted checkpoints fail closed.

Retention

Start with a dry run:

okto-nexus admin prune \
  --project-root /absolute/path/to/project \
  --dry-run

Then execute:

okto-nexus admin prune \
  --project-root /absolute/path/to/project \
  --messages-keep-days 30 \
  --vacuum

Retention spans the whole shared store even though --project-root is validated as the command anchor. --vacuum is the only option that compacts freed pages on disk. There is no always-running coordination reaper; auto_prune_on_start is a bounded opportunistic pass.

Issue legacy keys

okto-nexus admin issue-keys \
  --project-root /absolute/path/to/project

This is additive and idempotent. Existing keys are never rotated. Newly issued plaintext keys are printed once.

Replay export

okto-nexus admin export \
  --project-root /absolute/path/to/project \
  --trace-id trc_... \
  --output nexus-events.ndjson

The first line is a manifest; subsequent lines are raw events ordered by event_id. CLI export is operator-shell access and remains available even when the REST replay flag is off.

Ephemeral monitor token

An authenticated agent can call poll_token_issue, give only the returned nxsept_... token and base URL to a read-only helper, renew it before expiry, and revoke it on teardown. The raw token is returned only on issue/renew.

Data model and migrations

The current schema contains 34 tables:

Area	Tables
Core coordination	`schema_migrations`, `workspaces`, `agents`, `sessions`, `events`, `channels`, `messages`, `tasks`, `handoffs`, `artifacts`, `message_deliveries`
Settings/security/catalogs	`settings`, `permission_presets`, `tag_keys`, `tag_values`, `capability_names`, `ephemeral_poll_tokens`
Search and memory	`message_embeddings`, `memories`, `memory_embeddings`
Governance/workflows	`governance_policies`, `approvals`, `handoff_dependencies`, `policies`, `policy_versions`, `agent_policy_bindings`, `comm_presets`, `comm_preset_versions`, `agent_comm_binding`, `agent_groups`, `agent_group_members`, `guardrails`, `guardrail_versions`, `guardrail_assignments`

Not every table is workspace-scoped: agents and catalogs are global, inbox deliveries are keyed by recipient identity, and bindings/control-plane records have their own ownership rules.

Migrations are embedded in the package and applied in order:

001–008: core schema, close metadata, handoff payload/result, presence, durable inbox deliveries, leases, and session secrets;
009–015: API keys, settings, permissions, embeddings, tags/scopes, capability catalog, and trace IDs;
016–021: governance, approvals, verification, dependencies, memory, and health/event indexes;
022–026: versioned attachable policies, communication presets, display colors, groups/guardrails, and ephemeral poll tokens.

Nexus refuses to run against an unsupported newer schema. Runtime SQLite databases and their WAL/SHM/journal sidecars are ignored and must not be committed.

Errors

The domain/MCP contract has a closed catalog of 29 canonical codes:

Area	Codes
Workspace	`WORKSPACE_REQUIRED`, `WORKSPACE_UNRESOLVED`, `WORKSPACE_MISMATCH`
Validation/identity	`VALIDATION_ERROR`, `NOT_FOUND`, `NOT_OWNER`, `PERMISSION_DENIED`
Catalog/control plane	`TAG_IN_USE`, `CAPABILITY_IN_USE`, `POLICY_IN_USE`, `COMM_PRESET_IN_USE`
Governance	`POLICY_DENIED`, `QUOTA_EXCEEDED`, `GUARDRAIL_DENIED`, `CONFLICT`
Dependencies/state	`DEPENDENCY_NOT_FOUND`, `DEPENDENCY_NOT_MET`, `INVALID_TRANSITION`, `INVALID_STREAM`
Handoffs	`HANDOFF_ALREADY_CLAIMED`, `NOT_ELIGIBLE_TO_CLAIM`
Content/path	`CONTENT_TOO_LARGE`, `PATH_OUTSIDE_WORKSPACE`
Infrastructure	`CONFIG_ERROR`, `MIGRATION_ERROR`, `DB_ERROR`, `RENDER_ERROR`
Compatibility/internal	`MIGRATED`, `INTERNAL_ERROR`

REST adapters also use transport-specific codes such as AUTH_FAILED, CROSS_ORIGIN_BLOCKED, INVALID_PARAM, INVALID_WINDOW, INVALID_SETTING, EMBEDDINGS_UNAVAILABLE, and INTERNAL.

Development

uv sync --extra dev
uv run pytest -q

For the complete HTTP/embedding environment:

uv sync --extra serve --extra dev
uv run pytest -q

Frontend:

cd frontend
npm ci
npm run build

The build writes packaged static assets under src/okto_nexus/adapters/inbound/http/static/.

Release checks:

uv lock --check
uv build --out-dir dist/release-0.1.4
uvx twine check \
  dist/release-0.1.4/okto_nexus-0.1.4-py3-none-any.whl \
  dist/release-0.1.4/okto_nexus-0.1.4.tar.gz

Publish only explicitly named current-version artifacts. The top-level dist/ may contain older builds.

Project layout

src/okto_nexus/
  adapters/
    inbound/
      cli/                  serve, tail, admin
      http/                 FastAPI, REST, SSE, packaged SPA
      mcp/                  server, resources, projections, 43/46 tools
    outbound/
      sqlite/               repositories and migrations adapter
      embedding/            optional semantic provider
      file/ sharedmd/       artifact and derived-view I/O
      telemetry/ tokenizer/ metrics support
  application/              use cases and ports
  domain/                   entities, routing, state machines, policies
  migrations/               001 through 026
  testing/                  reusable test/replay harnesses
frontend/                    React dashboard source
tests/                       unit, contract, integration, replay tests
docs/design/                 architecture and design records
pyproject.toml               package metadata and extras
uv.lock                      reproducible dependency lock

Contributing

See CONTRIBUTING.md for the supported Python and frontend development environments, validation commands, and pull-request workflow.

Troubleshooting

`event_wait` returns immediately

Pass timeout_seconds > 0. Omitted, null, and zero are snapshots by design.

An agent misses broadcasts

Check that it opened a session in the correct resolved workspace and continues to heartbeat. Read-only event/inbox checks do not advance presence.

A direct peer is missing from discovery

Authenticated discovery is filtered by communication reachability. Inspect the caller's outbound and the peer's inbound communication scopes/tags in the dashboard.

Capability or tag targeting fails

Create the capability/tag value in Registry first. Catalog validation is fail-closed.

Memory tools are absent

Set feature_memory=true and restart/reconnect. Unlike live behavior flags, this flag changes MCP tool registration.

Semantic search is unavailable

Use embedding_mode=stub or install the embeddings/serve extra and use embedding_mode=local. off intentionally disables search.

`DB_ERROR` reports a lock

If details.retryable=true, retry the same call after the competing writer commits. Avoid opening the runtime SQLite file with tools that hold long write transactions.

The hub says another server already owns the home

serve holds {home}/nexus.serve.lock and refuses a second server using that same home. Stop the other hub or choose a different --home; changing only --db-path does not change the lock scope.

Workspace path is rejected

Pass an existing absolute path. Nexus canonicalizes the real path before deriving the workspace ID and validating artifact containment.

A monitor cannot mutate state

That is expected for nxsept_ tokens. They are intentionally read-only and accepted only on the four monitor endpoints.

Cached docs appear stale

Call nexus_info and compare surface_revision and resource_versions before reusing cached MCP reference content.

Security and limitations

Report vulnerabilities privately according to SECURITY.md. Do not include vulnerability details, credentials, tokens, or private workspace data in a public issue.

Nexus is designed for local or controlled single-tenant coordination, not as a public multi-tenant broker.
It does not terminate TLS. Put an authenticated TLS reverse proxy in front of a remote bind.
The dashboard shell and public health/info/license assets remain public; data/control REST requires authentication outside loopback.
API keys are hash-only at rest and shown once. Regeneration invalidates the old key immediately.
Session secrets are stored in plaintext in the local SQLite database; anyone who can read that file is inside the session trust boundary.
Channels are labels, not security boundaries.
SQLite is the only built-in coordination store. There is no Redis, PostgreSQL, or cloud broker adapter.
There is no always-running coordination scheduler/reaper. Expiry is checked opportunistically and retention runs manually or at startup when enabled.
The HTTP server does use worker threads/tasks for operational needs such as blocking waits and telemetry; “no scheduler” does not mean “no threads.”
Message durability is bounded by message retention and explicit reset.
Memory is experimental and changes the MCP surface at startup.
Artifact file references are confined to the canonical workspace.
Avoid committing runtime databases, sidecars, metrics output, and local secrets to version control.

Future direction

Likely extension points are additional durable-store adapters, a push-backed waiter that removes internal sleep polling, stronger remote deployment packaging, and further generated documentation from the live MCP schemas. The delivered surface already includes permissions, catalogs, communication scopes, policies, guardrails, HITL, trace correlation, verified/DAG handoffs, memory, health, replay, ephemeral poll tokens, embeddings, metrics, REST, SSE, and the dashboard.

Release notes

0.1.4 — current

Dashboard, observability, and coordination-guidance release. The MCP guidance contract advances to surface revision 32; the database schema is unchanged and the latest migration remains 026.

Added detailed and compact activity-based agent graph modes, profile colours, live status badges, richer relationship context, and graph conversations.
Added an event timeline, expanded event and handoff filters, message detail hydration, workspace display names, and catalog import/export workflows.
Improved dashboard views for agents, approvals, communication, events, guardrails, handoffs, messages, policies, and workspaces.
Extended observability APIs and repositories with message lookup, filtered handoff/event queries, and bucketed event timeline data.
Completed PyPI project URLs, keywords, and supported-Python classifiers, with a focused metadata regression test.
Added contributor and security policies, structured GitHub issue forms, and a dedicated documentation-assets location for product screenshots.
Clarified that agents follow the role and communication profile returned by agent_whoami unless the user supplies a task-scoped override, without bypassing platform governance.
Defined handoffs as the required, traceable mechanism for executable work, broadcasts as shared alignment or information fan-out, and direct messages as status, clarification, and informal coordination channels.
Rewrote this README against the current CLI, transport/authentication model, dashboard, 43/46-tool surface, seven routing strategies, message retention, governance features, verification/DAG workflows, 34-table schema, 29-code error catalog, operations, testing, and limitations.
Synchronized pyproject.toml and the root project entry in uv.lock at version 0.1.4, and aligned the package/dashboard license label with the addendum that is actually included in LICENSE.
Removed tracked runtime SQLite databases and added ignore coverage for database files and journal/WAL/SHM sidecars.
Verified release archives contain the dashboard and migrations without runtime SQLite databases or sidecars.

0.1.2

Documentation-accuracy sweep at surface revision 31.
Corrected pre-flight, monitoring, heartbeat, inbox, trace, policy, HITL, artifact-audience, and target-grammar documentation.
Kept 12 deep reference resources versioned and reduced the resident cuttable surface to about 26.3k characters.
Corrected the token-reduction gate to include experimental-surface growth.

0.1.1

Hardened authenticated self-only identity/session rules and permission checks.
Added permission-gated workspace paths, shared view, health, and memory.
Added configurable aggregate metrics telemetry.

0.1.0

Added ephemeral remote-monitor tokens and read-only monitor endpoints.
Added attachable policies, communication presets, guardrail/group administration, enriched agent graph cards, and loopback trust hardening.
Made memory a registration-time experimental MCP surface.

0.0.x

Built the MCP reference-resource system, HTTP hub, live dashboard, inbox receipts, semantic search, monitoring guidance, and dashboard observability waves.

License

Okto Nexus is distributed under the Elastic License 2.0 together with the project's SaaS, competing-service, internal-use, and branding addendum. It permits internal and qualifying single-tenant use and prohibits the specified multi-tenant, white-label/OEM, competing, and large-scale internal-platform uses; applicable notices and attribution remain required.

Read the complete LICENSE before use or redistribution. It is also included in the source distribution and served by the running hub at GET /api/v1/license.