AgentSpec Marketplace
AgentSpecs are structured prompt libraries that enhance a tulpa agent’s domain knowledge, behavioral patterns and response templates. Unlike extensions, AgentSpecs require no backing server — the spec itself is the complete deliverable. Install an AgentSpec and your agent immediately gains domain-specific guidance.
What AgentSpecs are
- A structured JSON format for community-authored prompt libraries
- A lightweight marketplace with content-quality review
- A composition model for layering multiple domain behaviors onto a single agent
- A way to make your agent better in specific professional domains without touching code
What AgentSpecs are not
- Not an extension. AgentSpecs have no endpoint, no permissions, no delegation tokens. They cannot take actions or read user data.
- Not a personality override. Your personality settings (verbosity, formality, proactivity) remain in control.
- Not fine-tuning. AgentSpecs are context-injected at prompt time. They are inspectable, removable and composable.
Content sections
Each spec contains up to five sections of domain knowledge:
| Section | Purpose | Max entries |
|---|---|---|
| Knowledge | Domain concepts the agent should understand | 20 |
| Behavioral rules | Context-triggered instructions for specific situations | 15 |
| Tool preferences | Guidance on which agent capabilities to prefer | 10 |
| Cadences | Follow-up timing guidance for domain-specific relationships | 10 |
| Templates | Response templates with formality variants | 10 |
Every entry has a stable key field used for deduplication and conflict resolution across specs.
Categories
| Category | Description |
|---|---|
sales | Enterprise sales, pipeline management, deal navigation |
recruiting | Talent sourcing, candidate engagement, hiring loops |
community | Community management, event organizing, membership |
consulting | Client relationship management, delivery follow-ups |
fundraising | Investor relations, pitch follow-ups, board management |
academia | Conference networking, collaboration, peer review |
job_search | Job hunting strategy, recruiter engagement, interview prep |
executive | Board relationships, stakeholder management |
creative | Creative industry networking, portfolio sharing |
general | Domain-agnostic behavioral enhancements |
Composition
You can install up to 10 specs but only 3 can be active at once. Active specs are layered by priority — when two specs define the same key in the same section, the higher-priority spec wins.
Default priority is installation order (most recent = highest). You can manually reorder at any time.
How specs interact with your data
- Memory wins over specs. If your agent has learned a preference that contradicts a spec, it follows your preference.
- Personality controls tone. Templates render the variant matching your formality setting.
- Your cadences take precedence. Spec-suggested follow-up timing is advisory — your existing cadence preferences override it.
Submission and review
Submitting a spec requires a verified tulpa handle. Specs go through automated guardrail checks and a content review before listing.
Automated checks reject specs that attempt to override safety rules, extract user data, impersonate entities or use manipulation tactics. Specs that pass automated checks enter human review evaluating guardrail safety, content accuracy, manipulation absence and identity integrity.
API
Browsing (public, no auth required)
- Search — query by keyword, filter by category
- Get — retrieve the full spec JSON
Managing (owner auth required)
- Install — fetch and store a listed spec locally
- Uninstall — remove a spec and re-normalize priorities
- Reorder — set priority ordering across all installed specs
- Toggle — activate or deactivate an installed spec
- Update — pull a newer marketplace version (priority and active state preserved)
No auto-updates — specs are never updated without user action.