Tessl operates across three interconnected surfaces.

The Tessl Registry is a public package manager for versioned, evaluated agent context. As of early 2026, it contains more than 10,000 tiles -- Tessl's term for curated context packages -- covering open-source libraries, APIs, and developer tooling, discoverable and installable via CLI or MCP server. The tiles distinction matters: a tile is not a code package. It is structured, agent-readable context about how to use a package correctly. The analogy Tessl uses is that tiles are to context what npm packages are to code.

Each tile carries three evaluation signals.

Quality scores assess whether the context is well-constructed: are descriptions clear enough for an agent to match the tile to the right user intent, are usage examples concrete, is error handling addressed. A tile scoring 40% on quality is not broken -- it will not get invoked reliably.

Impact scores measure whether the tile actually improves agent performance against real-world task scenarios. Tessl's benchmark data is specific: on the Vercel-published Next.js evaluation suite, top agents scored 42% without context; Tessl tiles lifted that to 92%. Across 270 libraries using LLM-generated evaluation tasks, average agent success rate moved from 60% to 81%, in less time. Context density degrades performance -- a key Tessl finding -- validating the importance of precision over volume: a 3KB security guidance file scored 85%; the 20KB version of the same guidance scored 81%.

The abstraction adherence data adds a different dimension. In a separate evaluation by Tessl's AI engineering team -- measuring not just task completion but whether agents correctly use library abstractions rather than reimplementing from scratch -- Tessl tiles produced 35% higher abstraction adherence on average across libraries. For newer libraries released in the last three years, the improvement reached 50%. The library age curve is telling: agents underperform on both very old and very new libraries. Old libraries suffer from confusing or outdated web representation; new libraries lack sufficient training data. Tessl tiles flatten that curve. A LangGraph case study focused specifically on features introduced after the model's training cutoff showed 90% improvement in agent performance on those post-cutoff features when tiles were present.

The evaluation methodology itself is worth understanding. Tessl compared three scenarios: baseline agent with no additional context; agent with access to the full package source code; and agent with Tessl tiles. Source code access improved performance, but at the cost of speed and token efficiency -- the agent had to manually search through unoptimized, high-volume context to extract what it needed. Tiles outperformed on both correctness and efficiency: curated, structured, agent-readable context that the agent can reach for precisely when needed.

Security scores, launched March 17, 2026, in partnership with Snyk, address the attack surface unique to agent context: natural language that shapes autonomous behavior. Snyk's ToxicSkills research scanned nearly 4,000 public skills and found 36% contained prompt injection techniques. Three lines of markdown in a skill file were sufficient to instruct an agent to read SSH keys and exfiltrate them. Standard scanners skip markdown files entirely. Snyk's engine, built on research from Invariant Labs, uses calibrated models combined with deterministic rules to analyze behavioral intent: prompt injection, obfuscated and base64-encoded variants, malware payloads, credential mishandling, toxic flow patterns, and suspicious external downloads. The integration spans the full lifecycle -- scan at publish, visible at browse, gate at install (CLI warns or blocks on high and critical findings), and feedback loop at authoring. Version pinning enforces that if the latest GitHub version of a skill differs from the last scanned registry version, the CLI installs at the last secure commit.

The three signals together answer one question: should I install this tile? Quality without impact means well-constructed context that does not make the agent better. Impact without security means flying blind on risk.

Tessl's internal architecture distinguishes between two types of context: knowledge and rules. Knowledge is the large, structured documentation that agents reach for dynamically when they need it -- a tile's library reference, linked from a small pointer in the agent's configuration. Rules are compact behavioral instructions loaded into context always -- short enough that the agent attends to every one. The distinction is not cosmetic. Podjarny's data shows that agents given 10 behavioral instructions follow each less carefully than agents given 3. Rules must be sparse; knowledge must be reachable. Tessl's tile structure enforces this separation by design.

The Tessl Framework is a spec-driven development workflow that prompts agents to capture intent in specifications before coding, keeping agent and developer aligned on what to build. It launched in private beta in September 2025.

The CLI (npx tessl) and MCP server are the primary developer-facing interfaces, handling search, install, and skill evaluation across agents.