diff --git a/src/library/type_context.h b/src/library/type_context.h
new file mode 100644
index 0000000000..ac18d9e30b
--- /dev/null
+++ b/src/library/type_context.h
@@ -0,0 +1,121 @@
+/*
+Copyright (c) 2016 Microsoft Corporation. All rights reserved.
+Released under Apache 2.0 license as described in the file LICENSE.
+
+Author: Leonardo de Moura
+*/
+#pragma once
+#include "library/metavar_context.h"
+
+namespace lean {
+enum class transparent_mode { ALL, SEMIREDUCIBLE, REDUCIBLE, NONE };
+
+/* \brief Cached information for type_context. */
+class type_context_cache {
+    typedef std::unordered_map<name, optional<declaration>, name_hash> transparent_cache;
+    typedef expr_struct_map<expr> infer_cache;
+    typedef expr_map<expr>        whnf_cache;
+    typedef expr_struct_map<optional<expr>> instance_cache;
+    typedef expr_struct_map<optional<expr>> subsingleton_cache;
+    environment                   m_env;
+    options                       m_options;
+    name_map<projection_info>     m_proj_info;
+
+    /* We only cache inferred types if the metavariable assignment was not accessed.
+       This restriction is sufficient to make sure the cached information can be reused
+       because local declarations have unique global names, and these names
+       are never reused. So, a term `t` containing locals `l_1, ..., l_n`
+       will have the same type in any valid local context containing `l_1, ..., l_n`.
+
+       \remark The inferred type does not depend on reducibility annotations. */
+    infer_cache                   m_infer_cache;
+
+    /* Cache for whnf (without delta). As in m_infer_cache, we only cache results if the
+       metavariable assignment was not used. */
+    whnf_cache                    m_whnf_cache;
+
+    /* Mapping from name to optional<declaration>, this mapping is faster than the one
+       at environment. Moreover, it takes into account constant reducibility annotations.
+       We have four different modes.
+       - ALL (everything is transparent).
+       - SEMIREDUCIBLE (semireducible and reducible constants are considered transparent).
+       - REDUCIBLE (only reducible constants are considered transparent).
+       - NONE (everything is opaque).
+
+       \remark In SEMIREDUCIBLE and REDUCIBLE modes, projections and theorems are considered
+       opaque independently of annotations. In ALL mode, projections are considered opaque.
+
+       The ALL mode is used for type inference where it is unacceptable to fail to infer a type.
+       The SEMIREDUCIBLE mode is used for scenarios where an is_def_eq is expected to succeed
+       (e.g., exact and apply tactics).
+       The REDUCIBLE mode (more restrictive) is used during search (e.g., type class resolution,
+       blast, etc).
+       The NONE mode is used when normalizing expressions without using delta-reduction. */
+    transparent_cache             m_transparent_cache[4];
+
+    /* We have two modes for caching type class instances.
+
+       In the default mode (m_frozen_mode == false), whenever a type_context object is
+       initialized with a local_context, we reset m_local_instances_initialized flag.
+       and store a copy of the initial local_context. Then, the first time type class resolution
+       is invoked, we collect local_instances, if they are equal to the ones in m_local_instances,
+       we do nothing. Otherwise, we reset m_local_instances with the new local_instances, and
+       reset the cache m_local_instances.
+
+       When frozen mode is set, we reset m_local_instances_initialized, the instance cache,
+       and the vector local_instances. Then, whenever a type_context object is created
+       (and debugging code is enabled) we store a copy of the initial local context.
+       Then, whenever type class resolution is invoked and m_local_instances_initialized is false,
+       we copy the set of frozen local_decls instances to m_local_instances.
+       If m_local_instances_initialized is true, and we are in debug mode, then
+       we check if the froze local_decls instances in the initial local context are indeed
+       equal to the ones in m_local_instances. If they are not, it is an assertion violation.
+
+       We use the same cache policy for m_subsingleton_cache. */
+    bool                          m_frozen_mode;
+    bool                          m_local_instances_initialized;
+    local_context                 m_init_local_context;
+    std::vector<pair<name, expr>> m_local_instances;
+    instance_cache                m_instance_cache;
+    subsingleton_cache            m_subsingleton_cache;
+
+    friend class type_context;
+    void init(local_context const & lctx);
+public:
+    type_context_cache(environment const & env, options const & opts);
+};
+
+class type_context : public abstract_type_context {
+    typedef type_context_cache cache;
+    local_context     m_lctx;
+    metavar_context & m_mctx;
+    cache *           m_cache;
+    bool              m_cache_owner;
+    transparent_mode  m_mode;
+public:
+    type_context(metavar_context & mctx, local_context const & lctx, type_context_cache & cache,
+                 transparent_mode m = transparent_mode::REDUCIBLE);
+    type_context(environment const & env, options const & opts, metavar_context & mctx, local_context const & lctx,
+                 transparent_mode m = transparent_mode::REDUCIBLE);
+    virtual ~type_context();
+
+    virtual environment const & env() const;
+    virtual expr whnf(expr const & e);
+    virtual expr infer(expr const & e);
+    virtual expr check(expr const & e);
+    virtual bool is_def_eq(expr const & e1, expr const & e2);
+
+    virtual expr relaxed_whnf(expr const & e);
+    virtual bool relaxed_is_def_eq(expr const & e1, expr const & e2);
+
+    virtual optional<expr> is_stuck(expr const &);
+    virtual name get_local_pp_name(expr const & e) const;
+
+    virtual bool on_is_def_eq_failure(expr const &, expr const &);
+
+    bool is_prop(expr const & e);
+
+    optional<expr> mk_class_instance(expr const & type);
+    optional<expr> mk_subsingleton_instance(expr const & type);
+};
+}